An Analysis of LIGO Glitches Using t-SNE During the First Part of the Fourth LIGO-Virgo-KAGRA Observing Run

arXiv:2512.03440 · astro-ph.IM, gr-qc · Submitted 2025-12-03 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics. The day in astrophysics: observations, models and what the data forces.

Vera: I'm Vera.

Jocelyn: And I'm Jocelyn, here with Subrahmanyan.

Vera: We're looking at the paper: "An Analysis of LIGO Glitches Using t-SNE During the First Part of the Fourth LIGO-Virgo-KAGRA Observing Run".

Paper summary: Jocelyn: So looking at the full scope of "An Analysis of LIGO Glitches Using t-SNE During the First Part of the Fourth LIGO-Virgo-KAGRA Observing Run," what's really important is how they use this unsupervised learning to sort out these noise artifacts.

Vera: They successfully uncovered two distinct glitch groups at Livingston that follow seasonal behavior related to microseismic ground motion.

Subrahmanyan: And at Hanford, the research indicated that broadband glitches in the twenty to fifty hertz range are associated with the electrostatic drive system.

Jocelyn: So it’s about using this t-SNE and clustering approach to find those patterns we can then try to connect back to real-world conditions like ground motion or detector operations.

Vera: And the paper points out that these findings give strong indications of the possible origin of these glitch groups by showing their temporal and environmental links.

Subrahmanyan: The method itself, using t-SNE followed by Agglomerative Clustering to determine optimal number of groups with a Silhouette Score, is a specific way to organize this high dimensional data.

Jocelyn: And the work suggests that the next step in this research would be to examine relevant detector logbook entries to correlate these glitch group behaviors with those environmental or instrumental conditions they found.

Conclusion: Vera: So we've been looking at how this paper uses t-distributed Stochastic Neighbor Embedding to sort out noise transients from LIGO data during that first part of the fourth observing run and now we're getting to the conclusion.

Jocelyn: The title itself is pretty descriptive, focusing on using t-SNE to examine glitch groups. I mean, it’s a technical tool being used for something very messy like noise glitches.

Subrahmanyan: And the authors are taking this from unclustered Omicron data across all Q values to create a high-dimensional representation before they project it into two dimensions where morphological similarities show up. That's a specific way to look at the glitch shapes.

Vera: Exactly, and what we see is that this method helps them identify these three main groups—L1, L2, and L3—and then track how they change over time during October through January in O4a.

Jocelyn: Tracking their temporal evolution is key because it connects the noise events to actual changes happening in the observatory environment. Like how Group L1 became dominant by November and January.

Subrahmanyan: From a theoretical side, this suggests that these glitch groups aren't just random noise; they're following seasonal variations of microseismic ground motion, which is something we need to consider when modeling astrophysical signals.

Vera: And at Hanford, the paper found something different—broadband glitches in the twenty to fifty hertz range are actually linked to the electrostatic drive system. That’s a real instrumental finding there.

Jocelyn: So it’s not just about what happened on Earth, but how these environmental factors and specific hardware systems create noise that we have to account for when looking for actual gravitational waves.

Subrahmanyan: The implication is that if we want to clean up the data and see the real signals, we need to understand these underlying noise sources tied to both the environment and our detector setup.

Vera: That's what it boils down to, figuring out what’s causing these specific glitch patterns so we can better isolate the physics of spacetime itself.

Jocelyn: It makes you wonder what kind of follow-up studies they might do next to really nail down the origin of those L1 and L3 groups.

Louisiana State University

astro-ph.IM, gr-qc

Submitted: 2025-12-03

Updated: 2025-12-07

Comments: 30 pages, 26 figures

Journal ref: APS Open Sci. 1, 000220 (2026)

DOI: 10.1103/nz94-2p7q

License: http://creativecommons.org/licenses/by/4.0/

The gist: An analysis of noise transients observed in LIGO data during the first part of the fourth observing run using t-distributed Stochastic Neighbor Embedding (t-SNE) to examine glitch groups is presented…

Key concepts

t-distributed Stochastic Neighbor Embedding (t-SNE)
A machine learning technique used to simplify complex data. It takes high-dimensional glitch features and maps them down into a simple two-dimensional space where glitches with similar shapes or characteristics are grouped together, making patterns easier to see.
Agglomerative Clustering
An unsupervised method applied after t-SNE. This process groups the now simplified data points (glitches) based on their proximity in the two-dimensional map. It assigns labels to these clusters, allowing researchers to define distinct glitch populations.
Glitch Groups (L1, L2, L3)
The study identified three main types of noise transients: Group L1 and L3 showed a strong link to seasonal ground motion at Livingston. Group L2 was prominent during low ground motion periods. These groups represent distinct physical origins or behaviors of the observed noise glitches.
Electrostatic Drive (ESD) System
A component within the LIGO detector system that uses an electrostatic drive voltage to control or influence noise. At Hanford, broadband glitches in the 20–50 Hz range were found to correlate with variations in this ESD system, suggesting it is a source for those specific noises.

Terminology

Summary

An analysis of noise transients observed in LIGO data during the first part of the fourth observing run using t-distributed Stochastic Neighbor Embedding (t-SNE) to examine glitch groups is presented

Methodology Overview

The study employs an unsupervised machine learning framework based on t-distributed Stochastic Neighbor Embedding (t-SNE) followed by Agglomerative Clustering to investigate the behavior of noise transients observed in LIGO data. This approach reduces the high-dimensional representation of glitches, built from unclustered Omicron data across all Q-values, into a two-dimensional space where morphological similarities emerge. p. 1

Where the paper says it

When excess power is identified in the data, Omicron characterizes it by analyzing its time-frequency tiles in spectrograms produced from the Q-transform [38]. This information is stored in unclustered files, and after preparing these data, we employ t-distributed Stochastic Neighbor Embedding (t-SNE) [39], a dimensionality reduction technique, to project the high-dimensional output of Omicron (which includes numerous tile-based features) into a two-dimensional space.

p. 1
Where the paper says it

When excess power is identified in the data, Omicron characterizes it by analyzing its time-frequency tiles in spectrograms produced from the Q-transform [38]. This information is stored in unclustered files, and after preparing these data, we employ t-distributed Stochastic Neighbor Embedding (t-SNE) [39], a dimensionality reduction technique, to project the high-dimensional output of Omicron (which includes numerous tile-based features) into a two-dimensional space.

The methodology involves projecting the high-dimensional output of Omicron (which includes numerous tile-based features) into a two-dimensional space using t-SNE. p. 1
Where the paper says it

When excess power is identified in the data, Omicron characterizes it by analyzing its time-frequency tiles in spectrograms produced from the Q-transform [38]. This information is stored in unclustered files, and after preparing these data, we employ t-distributed Stochastic Neighbor Embedding (t-SNE) [39], a dimensionality reduction technique, to project the high-dimensional output of Omicron (which includes numerous tile-based features) into a two-dimensional space.

Subsequently, Agglomerative Clustering is applied to group similar glitches and assign labels, with the Silhouette Score used to determine the optimal number of groups. p. 2
Where the paper says it

To facilitate this analysis, we first apply t-SNE to project the high-dimensional data into a two-dimensional space and then use Agglomerative Clustering [41] to group similar glitches and assign labels. The Silhouette Score [42] is used to evaluate clustering quality and determine the optimal number of groups.

Glitch Group Identification and Temporal Evolution

The analysis identified three main groups: Group L1 (purple), Group L2 (green), and Group L3 (orange). The temporal evolution of these groups was tracked by filtering the t-SNE output month by month during O4a. p. 6

Where the paper says it

An important advantage of this method is its ability to analyze the temporal evolution of glitches. Figure 6 presents the t-SNE output filtered month by month during O4a, providing an overview of which group was more prominent in each month.

Group L2 was most prominent during May, June, and July. p. 7
Where the paper says it

contribution of each group for the respective month. Comparing May, June, and July reveals a high similarity in cluster distribution, with Group L2 (green) being the most prominent during all three months.

In August, Group L1 began to exhibit higher densities. p. 7
Where the paper says it

The gray contours behind the data points indicate the estimated density, computed using Kernel Density Estimation (KDE) as implemented in Seaborn [49], where darker regions correspond to areas with higher data point concentrations. In August, Group L1 begins to exhibit higher densities, accounting for more than 50% of the data.

p. 7
Where the paper says it

The gray contours behind the data points indicate the estimated density, computed using Kernel Density Estimation (KDE) as implemented in Seaborn [49], where darker regions correspond to areas with higher data point concentrations. In August, Group L1 begins to exhibit higher densities, accounting for more than 50% of the data.

By November and January, Group L1 reached nearly 80% of the data. p. 7
Where the paper says it

In August, Group L1 begins to exhibit higher densities, accounting for more than 50% of the data. It slightly contracts in September (when an increase in the glitch rate from Group L3 is observed) but remains the dominant group in subsequent months, reaching nearly 80% of the data in November.

p. 7
Where the paper says it

In August, Group L1 begins to exhibit higher densities, accounting for more than 50% of the data. It slightly contracts in September (when an increase in the glitch rate from Group L3 is observed) but remains the dominant group in subsequent months, reaching nearly 80% of the data in November.

Correlation with Environmental Conditions

At Livingston, the most prominent glitch populations during O4a were closely correlated with seasonal variations in ground motion, particularly within the 0.1–1 Hz frequency range. p. 3

Where the paper says it

This shows the range where we observed the most common glitches at LLO during O4a, which occurred between 10 and 100 Hz. The increase in glitch occurrence at lower frequencies is associated with the improved sensitivity of the interferometer in this frequency range [7].

The two groups characterized by arch-like spectrogram morphologies (Group L1 and Group L3) were dominant during months of elevated ground motion. p. 4
Where the paper says it

One class of these transients, referred to as Scattered Light or slow scattering [43, 45], was characterized by arch-like features in their spectrograms. Given this, and based on the seasonal pattern in which ground motion in specific frequency bands tends to exhibit higher amplitudes during the winter months, we identified a potential correlation between such low-frequency glitches observed during O4a and elevated ground motion levels [46].

Conversely, Group L2 stood out during periods of low ground motion and showed no significant dependence on any of the seismic frequency bands tested. p. 2
Where the paper says it

Section 2 presents the glitch population observed at LLO during O4a, together with an analysis of ground motion across different frequency bands. The section also evaluates the correlation between ground motion and glitch rate, highlighting how the low-frequency glitches impacted the interferometer during the analyzed observing period.

Detector Specific Findings at Hanford

At Hanford (LHO), a different glitch landscape was observed. The detector experienced a high occurrence of broadband glitches in the 20–50 Hz range, particularly at lower SNR. p. 3

Where the paper says it

While some high-amplitude glitches between 10 and 13 Hz were detectable during O3, most of the lower-frequency glitches observed during O4a would not have been “visible” during O3, given the sensitivity limitations of the interferometer at that time. Figure 3 presents the average number of glitches per hour for each month during O4a, represented by the dark blue bars, considering glitches in the 10 to 2048 Hz range with a minimum SNR of 7.

These glitches were shown to be associated with the electrostatic drive (ESD) system, as indicated by their correlation with the RMS of the applied ESD voltage. p. 2
Where the paper says it

The section also evaluates the correlation between ground motion and glitch rate, highlighting how the low-frequency glitches impacted the interferometer during the analyzed observing period. A parallel analysis for LHO is provided in Section 3, focusing on the evolution of broadband glitches during O4a, which were primarily associated with internal instrumentation couplings rather than environmental factors.

Conclusions and Future Work

The analysis successfully uncovered two glitch groups at LLO that follow seasonal behavior of microseismic ground motion. At LHO, the research indicated that broadband glitches in the 20–50 Hz range are associated with the ESD system. p. 2

Where the paper says it

The section also evaluates the correlation between ground motion and glitch rate, highlighting how the low-frequency glitches impacted the interferometer during the analyzed observing period. A parallel analysis for LHO is provided in Section 3, focusing on the evolution of broadband glitches during O4a, which were primarily associated with internal instrumentation couplings rather than environmental factors.

p. 2
Where the paper says it

The section also evaluates the correlation between ground motion and glitch rate, highlighting how the low-frequency glitches impacted the interferometer during the analyzed observing period. A parallel analysis for LHO is provided in Section 3, focusing on the evolution of broadband glitches during O4a, which were primarily associated with internal instrumentation couplings rather than environmental factors.

The technique aims to uncover monthly behavior and correlate glitch groups with environmental or instrumental conditions by examining relevant detector logbook entries. p. 2
Where the paper says it

The Silhouette Score [42] is used to evaluate clustering quality and determine the optimal number of groups. These group labels serve primarily as a means to organize and visualize the temporal behavior of distinct glitch populations, to support their possible association with environmental or instrumental conditions, and to enable the classification of previously unseen data.

The findings provide strong indications of the possible origin of these glitch groups. p. 1
Where the paper says it

The main motivation behind these efforts is to support the overarching goal of glitch studies: identifying their physical origin to eliminate or mitigate their impact on gravitational-wave detection. This work presents an analysis of the main glitch groups observed during O4a, investigating their potential correlations with environmental and/or instrumental variations using the methodology detailed in [35].

p. 1
Where the paper says it

The main motivation behind these efforts is to support the overarching goal of glitch studies: identifying their physical origin to eliminate or mitigate their impact on gravitational-wave detection. This work presents an analysis of the main glitch groups observed during O4a, investigating their potential correlations with environmental and/or instrumental variations using the methodology detailed in [35].

Acknowledgments

The authors thank Elenna Capote and Jess McIver for their very useful insights, comments, and suggestions. This work was supported by the National Science Foundation under grant numbers PHY-2110509 and PHY-2409740. The material is based upon work supported by NSF’s LIGO Laboratory, a major facility fully funded by the National Science Foundation. p. 1

Where the paper says it

One example is Gravity Spy, an interdisciplinary project that integrates machine learning, citizen science, and input from LIGO scientists to classify glitches [31, 32]. Glitches are named based on their morphology and/or known or suspected physical origin.

Additionally, the work utilizes the LIGO computing clusters and data from the Advanced LIGO detectors; the authors are grateful for the computational resources provided by the LIGO Laboratory. p. 1
Where the paper says it

This work presents an analysis of the main glitch groups observed during O4a, investigating their potential correlations with environmental and/or instrumental variations using the methodology detailed in [35]. This investigation identifies patterns among glitches using the information provided by Omicron [36, 37], the primary tool in LIGO used to detect noise transients in detector data.

Appendix A Details

The appendix presents a comparison of measurements from seismometers located near the test masses at LIGO Livingston. The analysis observed that in general, motion in the three directions follows similar trends, suggesting that any directional component can be informative when analyzing behavioral variations. p. 9

Where the paper says it

Given the previous insights from the t-SNE output, we investigate possible relationships between the identified groups and seismic noise. In general, amplitude variations in ground motion across the three directions exhibit similar overall trends.

For the LMS and HMS bands, motion in the ˆx and ˆy directions is similar and generally larger than in the ˆz direction. p. 4
Where the paper says it

Ground motion at LIGO is measured by seismometers positioned at various locations around the interferometer, which record ground velocity, typically expressed in nanometers per second (nm/s). Measurements are taken along three axes: ˆx, aligned with the input laser, ˆy, perpendicular to ˆx; and ˆz, representing vertical motion orthogonal to the x–y plane.

Conversely, in the anthropogenic band, the ˆz direction exhibits higher amplitudes and larger variations. p. 5
Where the paper says it

3 and 1 Hz, also measured in the ˆx direction by the ETMY seismometer. • Anthropogenic band: ground motion between 1 and 3 Hz, measured in the ˆz direction by the ETMY seismometer.

The ETMY seismometer showed higher median values and larger variations than ETMX in the anthropogenic band, likely due to its closer proximity to nearby roads. p. 9
Where the paper says it

Appendix A presents a study of seismometers installed at different locations, illustrating this behavior and providing pairwise comparisons between them. The ETMY location exhibits a higher amplitude in the anthropogenic band, which led to its seismometer being chosen as the reference in this analysis — although any of the seismometers could have been selected.

Table A1 Data Summary

Table A1 summarizes the seismometer channels and corresponding position, arm, and motion directions. The data shows that for LMS and HMS bands, motion in the ˆx and ˆy directions is similar and generally larger than in the ˆz direction. p. 4

Where the paper says it

Ground motion at LIGO is measured by seismometers positioned at various locations around the interferometer, which record ground velocity, typically expressed in nanometers per second (nm/s). Measurements are taken along three axes: ˆx, aligned with the input laser, ˆy, perpendicular to ˆx; and ˆz, representing vertical motion orthogonal to the x–y plane.

In contrast, in the anthropogenic band, the ˆz direction exhibits higher amplitudes and larger variations. p. 5
Where the paper says it

3 and 1 Hz, also measured in the ˆx direction by the ETMY seismometer. • Anthropogenic band: ground motion between 1 and 3 Hz, measured in the ˆz direction by the ETMY seismometer.

The ETMY seismometer shows higher median values and larger variations than ETMX in the anthropogenic band, likely due to its closer proximity to nearby roads. p. 9
Where the paper says it

Appendix A presents a study of seismometers installed at different locations, illustrating this behavior and providing pairwise comparisons between them. The ETMY location exhibits a higher amplitude in the anthropogenic band, which led to its seismometer being chosen as the reference in this analysis — although any of the seismometers could have been selected.

Figure 12 Analysis

The analysis of glitch rates as a function of ground motion amplitude reveals that the glitch rates for Group L1 (purple) and Group L3 (orange) exhibit a positive correlation with increasing amplitudes in the LMS and HMS bands. In contrast, this trend is not observed for the anthropogenic band. p. 4

Where the paper says it

We analyze the Root Mean Square (RMS) of the ground motion within specific frequency ranges, known as Band-Limited RMS (BLRMS) [47]. The frequency bands considered correspond to the microseismic band-usually associated with ocean waves-and the anthropogenic band, which is generally linked to human activities such as passing trains or nearby vehicular traffic.

The analysis confirmed a strong correlation between Groups L1 and L3 and both the LMS and HMS bands. p. 2
Where the paper says it

Section 2 presents the glitch population observed at LLO during O4a, together with an analysis of ground motion across different frequency bands. The section also evaluates the correlation between ground motion and glitch rate, highlighting how the low-frequency glitches impacted the interferometer during the analyzed observing period.

Table 1 Correlation Summary

Table 1 summarizes the Pearson correlation coefficients for each case, computed from more than 2,000 valid data points. Overall, the results confirm a strong correlation between Groups L1 and Group L3 and both the LMS and HMS bands. p. 7

Where the paper says it

In December, for instance, the low density of points in this region results in an almost complete absence of KDE contours. This does not imply the disappearance of Group L2, but rather that it was no longer among the most significant groups.

Group L2 exhibits a weak correlation with the LMS band. p. 5
Where the paper says it

Light blue bars with hatching indicate the subset of glitches with frequencies below 25 Hz. • Lower MicroSeismic (LMS) band: ground motion between 0.

None of the groups shows a significant correlation with the anthropogenic band. p. 4
Where the paper says it

We analyze the Root Mean Square (RMS) of the ground motion within specific frequency ranges, known as Band-Limited RMS (BLRMS) [47]. The frequency bands considered correspond to the microseismic band-usually associated with ocean waves-and the anthropogenic band, which is generally linked to human activities such as passing trains or nearby vehicular traffic.

The corresponding correlation coefficients are very close to zero when considering only ground-motion amplitudes below 600 nm/s. p. 4
Where the paper says it

One class of these transients, referred to as Scattered Light or slow scattering [43, 45], was characterized by arch-like features in their spectrograms. Given this, and based on the seasonal pattern in which ground motion in specific frequency bands tends to exhibit higher amplitudes during the winter months, we identified a potential correlation between such low-frequency glitches observed during O4a and elevated ground motion levels [46].

Figure 18 ESD Correlation

Figure 18 illustrates that a clear correlation can be observed between variations in the ESD RMS signal and fluctuations in the glitch rate. The analysis confirmed a lack of correlation for Group H4 (r = 0. p. 2

Where the paper says it

Section 2 presents the glitch population observed at LLO during O4a, together with an analysis of ground motion across different frequency bands. The section also evaluates the correlation between ground motion and glitch rate, highlighting how the low-frequency glitches impacted the interferometer during the analyzed observing period.

p. 2
Where the paper says it

Section 2 presents the glitch population observed at LLO during O4a, together with an analysis of ground motion across different frequency bands. The section also evaluates the correlation between ground motion and glitch rate, highlighting how the low-frequency glitches impacted the interferometer during the analyzed observing period.

p. 2
Where the paper says it

Section 2 presents the glitch population observed at LLO during O4a, together with an analysis of ground motion across different frequency bands. The section also evaluates the correlation between ground motion and glitch rate, highlighting how the low-frequency glitches impacted the interferometer during the analyzed observing period.

12) and a strong positive correlation for Group H1 with r = 0. p. 2
Where the paper says it

Section 2 presents the glitch population observed at LLO during O4a, together with an analysis of ground motion across different frequency bands. The section also evaluates the correlation between ground motion and glitch rate, highlighting how the low-frequency glitches impacted the interferometer during the analyzed observing period.

p. 2
Where the paper says it

Section 2 presents the glitch population observed at LLO during O4a, together with an analysis of ground motion across different frequency bands. The section also evaluates the correlation between ground motion and glitch rate, highlighting how the low-frequency glitches impacted the interferometer during the analyzed observing period.

p. 2
Where the paper says it

Section 2 presents the glitch population observed at LLO during O4a, together with an analysis of ground motion across different frequency bands. The section also evaluates the correlation between ground motion and glitch rate, highlighting how the low-frequency glitches impacted the interferometer during the analyzed observing period.

76.

Figure 21 Classification Correlation

Figure 21b shows a scatter plot of the hourly glitch rate for all O4a data classified as Group H4 versus the median ESD over the same periods, revealing no apparent correlation. In contrast, the same type of plot for all glitches classified as H1 illustrates a clear linear correlation in Figure 21a. p. 8

Where the paper says it

3% Figure 6: A scatter plot of the same t-SNE coordinates from LLO, filtered by month for each identified class. Each panel illustrates the behavior of the clusters during O4a, from May 2023 to January 2024.

The analysis confirmed a lack of correlation for Group H4 (r = 0.12) and a strong positive correlation for Group H1 with r = 0.76.

Figure 19 Morphological Differences

The analysis of the temporal occurrence suggests that, except for Group H4, all the main glitches are likely associated with the same underlying source. Their differences lie primarily in the recurrence and morphology of specific features: vertical “blips” (Group H1), shortduration horizontal lines (Group H2), and longer-duration lines that vary in frequency, as well as in the number and timing of repetitions (Group H3). p. 7

Where the paper says it

Some of these glitches exhibit longer durations, others show repetitions over time, and some vary in the number of harmonics. Despite occurring at similar frequencies, these morphological differences explain the formation of miniclusters within the group (see Figure 5).

The decrease in the number of glitches in Group H1 during August and September — associated with the implementation of the new controller — contrasts with the increase in Group H2 during the same period. p. 7
Where the paper says it

In August, Group L1 begins to exhibit higher densities, accounting for more than 50% of the data. It slightly contracts in September (when an increase in the glitch rate from Group L3 is observed) but remains the dominant group in subsequent months, reaching nearly 80% of the data in November.

Figure 7 and Figure 8 Morphology

Group L1 is dominated by glitches with arch-like structures at a frequency of approximately 12 Hz. p. 7

Where the paper says it

This does not imply the disappearance of Group L2, but rather that it was no longer among the most significant groups. By examining the spectrograms, we observe that Group L1 is dominated by glitches with arch-like structures at a frequency of approximately 12 Hz; examples are shown in Figure 7.

Group L3 exhibits arch-like patterns with a primary frequency around 20 Hz. p. 8
Where the paper says it

The value of n indicates the total number of glitches in that month, and the percentages represent the relative fraction of glitches assigned to each group based on the t-SNE output. (arch-like patterns) are observed in Group L3, but with a primary frequency around 20 Hz; examples of their spectrograms are presented in Figure 8.

Group L2 primarily consists of glitch classes identified by Gravity Spy as Extremely Loud, Blip, Koi Fish, Tomte, and others. p. 8
Where the paper says it

(arch-like patterns) are observed in Group L3, but with a primary frequency around 20 Hz; examples of their spectrograms are presented in Figure 8. Group L2 primarily consists of glitch classes identified by Gravity Spy [50] as Extremely Loud, Blip, Koi Fish, Tomte, and others (examples are given in Figure 9).

Figure 11 Ground Motion Trends

The temporal evolution reinforces the weak relationship between Group L2 and ground motion. This trend aligns with the emergence and persistence of Groups L1 and L3 in the t-SNE output, which show a strong correlation with increased ground motion. p. 9

Where the paper says it

Given the previous insights from the t-SNE output, we investigate possible relationships between the identified groups and seismic noise. In general, amplitude variations in ground motion across the three directions exhibit similar overall trends.

Figure 6 Temporal Distribution

The monthly distribution shows that Group L2 was most prominent during May, June, and July. In August, Group L1 began to exhibit higher densities. By November and January, Group L1 reached nearly 80% of the data.

Figure 20 Classification Distribution

Figure 20 shows that only Group H4 deviates from the overall trend of broadband glitches shown in Figure 18. The y-axis scales differ across panels, highlighting the predominance of Group H1, followed by Group H2. p. 7

Where the paper says it

It slightly contracts in September (when an increase in the glitch rate from Group L3 is observed) but remains the dominant group in subsequent months, reaching nearly 80% of the data in November. Group L3 fluctuates over time but remains consistently present, with its highest predominance occurring in September.

Figure 17 LHO Evolution

The t-SNE plots for LHO show that in May, less than one week of data was available, during which Group H4 dominated the distribution. In June, the bottom-center region of the t-SNE map becomes increasingly prominent. p. 7

Where the paper says it

contribution of each group for the respective month. Comparing May, June, and July reveals a high similarity in cluster distribution, with Group L2 (green) being the most prominent during all three months.

From October through January, the large cluster H1 and H2 became the most persistent feature in the t-SNE projection, collectively accounting for at least 70% of the observed glitches.

Figure 18 ESD RMS Correlation

The impact of the new alignment control system is evident at the beginning of August, with a significant drop in both ESD RMS and glitch rate. Another key moment occurs in October, when both show a substantial increase, related to the reallocation of the LSC feedforward from ETMX to ETMY.

Figure 19 Spectrogram Examples

The spectrograms for Group H1 exhibit frequencies around 12 Hz. The spectrograms for Group H3 exhibit frequencies around 20 Hz. p. 9

Where the paper says it

Figure 7: Examples of spectrograms present in Group L1. The spectrograms exhibit frequencies around 12 Hz, with variations in morphology.

Figure 7 and Figure 8 Spectrogram Examples

Group L3 exhibits arch-like patterns with a primary frequency around 20 Hz. Group L1 is dominated by glitches with arch-like structures at a frequency of approximately 12 Hz.

Figure 14 Detector Comparison

During low microseismic activity (Figure 14a), very few glitches are observed at either observatory, though some broadband glitches remain present at 15. In contrast, during periods of elevated microseismic motion (Figure 14b), both observatories exhibit a higher occurrence of transient noise spanning a broader frequency range. p. 4

Where the paper says it

Given this, and based on the seasonal pattern in which ground motion in specific frequency bands tends to exhibit higher amplitudes during the winter months, we identified a potential correlation between such low-frequency glitches observed during O4a and elevated ground motion levels [46]. This behavior is further supported by the ground motion data shown later in Figure 11.

Figure 13 LHO Frequency Range

The histogram highlights the most common glitch range at LHO during O4a, between 20 and 40 Hz.

Figure 16 LHO Classification

The classified output for LHO highlights four main groups, labeled Group H1 through Group H4.

Figure 22 Final Observation

Dedicated testing by the LHO team led to the implementation of a new feedback control configuration, which significantly reduced the noise and resulted in a more stable and stationary system by lowering the ESD drive voltage. This improved configuration was adopted for the O4b observing run, which began with no recurrence of these glitches. p. 0

Where the paper says it

An Analysis of LIGO Glitches Using t-SNE During the First Part of the Fourth LIGO-Virgo-KAGRA Observing Run Tabata Aira Ferreira, Gabriela Gonz´alez, Osvaldo Salas Louisiana State University, Baton Rouge, LA 70803, USA E-mail: tferreira@lsu. edu Abstract.

Figure 18 Bicoherence

A bicoherence analysis revealed that these broadband glitches were related to the electrostatic drive (ESD) system in the detector. The analysis confirmed a lack of correlation for Group H4 (r = 0.12) and a strong positive correlation for Group H1 with r = 0.76.

Conclusion Summary

The analysis successfully uncovered two glitch groups at LLO that follow seasonal behavior of microseismic ground motion. At LHO, the research indicated that broadband glitches in the 20–50 Hz range are associated with the ESD system. The findings provide strong indications of the possible origin of these glitch groups.

Final Synthesis

The study employed a fully unsupervised machine learning framework based on t-distributed Stochastic Neighbor Embedding (t-SNE) followed by Agglomerative Clustering to investigate the behavior of noise transients observed in LIGO data. This approach reduces the high-dimensional representation of glitches, built from unclustered Omicron data across all Q-values, into a two-dimensional space where morphological similarities emerge. The study employed this methodology independently to identify dominant glitch populations and track their evolution over time. p. 2

Where the paper says it

checking the clusters obtained from t-SNE with Gravity Spy labels demonstrates that, when applied to unclustered Omicron information, t-SNE produces groups that closely correspond to known glitch classes. In this work, the primary objective is to track the temporal evolution of glitches during the O4a observing run.

The results show that at Livingston, groups correlated with seasonal ground motion were dominant during autumn and winter. p. 0
Where the paper says it

We then track these groups over time and investigate correlations between their occurrence and environmental or instrumental conditions. At the Livingston observatory, the most common glitches during O4a were seasonal and associated with ground motion, whereas at Hanford, the most prevalent glitches were related to instrumental conditions.

At Hanford, broadband glitches were associated with the ESD system. p. 2
Where the paper says it

The section also evaluates the correlation between ground motion and glitch rate, highlighting how the low-frequency glitches impacted the interferometer during the analyzed observing period. A parallel analysis for LHO is provided in Section 3, focusing on the evolution of broadband glitches during O4a, which were primarily associated with internal instrumentation couplings rather than environmental factors.

The technique allows for correlating glitch group behavior with environmental or instrumental conditions by examining monthly trends. p. 2
Where the paper says it

The Silhouette Score [42] is used to evaluate clustering quality and determine the optimal number of groups. These group labels serve primarily as a means to organize and visualize the temporal behavior of distinct glitch populations, to support their possible association with environmental or instrumental conditions, and to enable the classification of previously unseen data.

The ideal next step is to examine relevant detector logbook entries to correlate the behavior of glitch groups with those modifications. p. 0
Where the paper says it

edu Abstract. This paper presents an analysis of noise transients observed in LIGO data during the first part of the fourth observing run, using the unsupervised machine learning technique t-distributed Stochastic Neighbor Embedding (t-SNE) to examine the behavior of glitch groups.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Unsupervised glitch grouping using t-SNE allows for the identification of patterns among glitches using the information provided by Omicron, which enables researchers to uncover patterns without prior knowledge or training data.

  2. Temporal evolution tracking enables analysis of glitch populations over time, as seen in Figure 6, showing which group was more prominent in each month, allowing for the organization and visualization of temporal behavior.

  3. Correlation analysis between glitch groups and environmental conditions can be performed by examining relationships with ground motion bands; for instance, the study found that Groups L1 and L3 exhibited a strong correlation with increased ground motion in the LMS and HMS bands.

  4. Detector-specific analysis allows for distinct noise source identification; at Hanford, the analysis revealed four main groups, labeled Group H1 through Group H4, where three were related to the ESD, highlighting a separation of instrumental origins.

  5. The system can correlate glitch behavior with specific instrumental interventions by analyzing logbook entries; this allows researchers to associate certain glitch groups with environmental or instrumental conditions and document their possible origin.

Abstract

This paper presents an analysis of noise transients observed in LIGO data during the first part of the fourth observing run, using the unsupervised machine learning technique t-distributed Stochastic Neighbor Embedding (t-SNE) to examine the behavior of glitch groups. Based on the t-SNE output, we apply Agglomerative Clustering in combination with the Silhouette Score to determine the optimal number of groups. We then track these groups over time and investigate correlations between their occurrence and environmental or instrumental conditions. At the Livingston observatory, the most common glitches during O4a were seasonal and associated with ground motion, whereas at Hanford, the most prevalent glitches were related to instrumental conditions.

Sources

Related papers