Towards an anomaly detection pipeline for gravitational waves at the Einstein Telescope Journal: Physics Letters B

Gianluca Inguglia, Huw Haigh, Kristyna Vitulova, Ulyana Dupletsa · 2025

Read Towards an anomaly detection pipeline for gravitational waves at the Einstein Telescope Journal: Physics Letters B by Gianluca Inguglia, Huw Haigh, Kristyna Vitulova, Ulyana Dupletsa (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2511-13154.txt.

Introduction

Gianluca Inguglia and three co-authors hunt gravitational waves by teaching a network what normal noise looks like. Their target is short bursts, under about two seconds, such as mergers that form intermediate-mass black holes, which are hard to separate from detector noise but brief enough for cheap machine learning. A deep convolutional autoencoder learns to reconstruct ordinary time-frequency spectrograms; anything it reconstructs badly is flagged as an anomaly. Tested on Einstein Telescope mock data, the initial system caught twenty-three percent of injected intermediate-mass merger signals in a single detector.

Adding weak supervision transformed the results: the model recovered every injected merger of that class. Regardless of mass or signal strength, at about four and a half false alarms per year for one interferometer. The pipeline flags anomalies without yet classifying them, a deliberate first stage.

Catalog record

CREATORGianluca Inguglia, Huw Haigh, Kristyna Vitulova, Ulyana Dupletsa
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2511-13154

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Towards an anomaly detection pipeline for gravitational waves at the Einstein Telescope Journal: Physics Letters B Gianluca Inguglia, Huw Haigh, Kristyna Vitulova, Ulyana Dupletsa Abstract We present the implementation of an anomaly–detection algorithm based on a deep convolutional autoencoder for the search for gravitational waves (GWs) in time-frequency spectrograms. Our method targets short-duration ( ≲ 2 ​ s \lesssim 2\,\text{s} ) GW signals, exemplified by mergers of compact objects forming or involving an intermediate-mass black hole (IMBH). Such short signals are difficult to distinguish from background noise; yet their brevity makes them well-suited to machine-learning analyses with modest computational requirements. Using the data from the Einstein Telescope Mock Data Challenge as a benchmark, we demonstrate that the approach can successfully flag GW-like transients as anomalies in interferometer data of a single detector, achieving an initial detection efficiency of 23% for injected signals corresponding to IMBH–forming mergers. After introducing weak supervision, the model exhibits excellent generalisation and recovers all injected IMBH–forming mergers, independent of their total mass or signal–to–noise ratio, with a false–alarm rate due to statistical noise fluctuations of approximately 4.5 events per year for a single interferometer operating with a 100% duty cycle. The method also successfully identifies lower–mass mergers leading to the formation of black holes with mass larger than ≃ 20 ​ M ⊙ \simeq 20\,M_{\odot} . Our pipeline does not yet classify anomalies, distinguishing between actual GW signals and noise artefacts; however, it highlights any deviation from the learned background noise distribution for further scrutiny. These results demonstrate that anomaly detection offers a powerful, model–independent framework for future GW searches, paving the way toward fully automated and adaptive analysis pipelines. 1 Introduction The LIGO-Virgo-KAGRA (LVK) Collaboration currently employs both modeled and unmodeled search techniques for gravitational-wave (GW) detection (see 2 for a recent overview), with a preference for modeled searches based on matched filtering. In this approach, precomputed waveform templates representing compact binary coalescences (CBCs) form a template bank that is cross-correlated with the interferometer data. When a template closely matches a potential signal, a trigger is generated, indicating a candidate GW transient for further inspection and validation. Although highly effective in signal detection, matched filtering is limited by the size and computational cost of the template bank. Since only a limited number of waveform templates can be included in the template bank, mismatches - a measure of how well a template approximates a true signal - are inevitable. It is standard practice to accept mismatches of up to 3% between a real signal and a template ( 41 ) , corresponding to a reduction in the signal-to-noise ratio (SNR) of up to 1.5%. This seemingly small SNR loss could be critical, especially for signals near the detection threshold, which might be missed. Unmodeled searches, on the other hand, aim to reconstruct GW signals without relying on specific waveform models. They target a wide variety of sources, and especially short, burst-like signals, which are the most difficult to distinguish from background noise and instrumental artefacts. Such short-duration, hard-to-model signals are ideal targets for machine learning algorithms, which can learn statistical patterns and representations directly from data. Machine and deep learning provide a powerful framework for classification, regression, and inference tasks, and are now widely adopted across most scientific domains. In deep learning, this is typically achieved through multi-layer neural networks trained via gradient-based optimisation to minimise task-specific loss functions ( 18 ; 9 ; 21 ; 43 ; 42 ) . Exploring machine learning approaches is becoming increasingly crucial in light of third-generation GW interferometers such as the Einstein Telescope (ET) ( 37 ; 26 ; 16 ; 4 ) or Cosmic Explorer (CE) ( 40 ; 20 ; 24 ) , representing the European and the United States proposals, respectively, for future ground-based GW detectors. Their improved sensitivity across a broader frequency range will lead to a substantial increase in detectable events, resulting in higher computational demands, not only for simulated data generation but also for detection and analysis of real data. This motivates the development of complementary or alternative search methods to enhance the discovery potential of current and future GW observatories. We reformulate GW searches as a general anomaly-detection problem: instead of matching data to predefined templates, the algorithm learns the structure of detector noise in frequency–time spectrograms and flags significant deviations as potential signals. In this work, we present the implementation and evaluation of an anomaly-detection algorithm for searching GW transients, based on convolutional autoencoders operating on spectrogram data from a single ET interferometer 1 1 1 As detailed in Sec. 4 , for single ET interferometer, we mean that we consider only one out of the 3 detector channels . We focus on burst signals, short GW transients that are produced, for instance, by the mergers of two black holes (BHs) involving or resulting in the formation of an intermediate-mass black hole (IMBH), a class of BHs with masses in the range 10^{2}-10^{5} M ⊙ ( 23 ) . We then demonstrate that our model is scalable and applicable across a broader range of BH masses. The paper is structured as follows. In Sec. 2 we introduce the science case of interest. We detail the network architecture and definition of a simple loss algorithm in Sec. 3 . In Sec. 4 , we then evaluate the ability of the model to reconstruct ET noise-only spectrograms and define the criteria by which observed data may be flagged as anomalous. We then investigate the performance of the model using the first Einstein Telescope Mock Data Challenge (MDC) ( 39 ; 38 ) dataset, focusing specifically on BBH mergers where the sum of source frame masses is greater than 100 M ⊙ . We conclude our discussion in Sec. 5 , highlighting future directions of this work. 2 Intermediate-mass black holes and their gravitational wave signatures Up to date, the LVK Collaboration has observed 218 GW candidate events, mostly associated with the merger of two black holes, as released in the latest GWTC-4.0 catalog ( 3 ) . Remarkably, several of these mergers involved IMBHs either in the merging components, as in the case of GW190521 ( 6 ; 7 ) or in the remnant, as for GW231123 ( 1 ) . One critical aspect of the GWs emitted during the formation of an IMBH, or during a merger involving IMBHs, is that the characteristic strain amplitude, for this type of transients, is distributed towards lower frequencies, well below 100 Hz, decreasing when the redshifted mass, M_{z}=(1+z)M(=m_{1}+m_{2}) , increases. Figure 1 shows the evolution of the merger frequency, exemplified by the f_{220} mode, for several total source-frame masses as a function of the redshift, indicating that the heavier and the further away the events, the lower the merger frequency, reaching values below the sensitivity level of next-generation ground-based interferometers such as ET. Figure 1: Theoretical frequency f_{220} as a function of redshift for binary black hole mergers assuming different total source-frame masses ( 14 ) , calculated using \alpha(a_{f})\simeq 0.536 and \varepsilon=0.07 ( 27 ; 25 ) . See A for further details. The red points indicate IMBH candidates detected via GWs, including GW190521 ( 6 ; 7 ) and GW231123 ( 1 )

Read the complete text