GlitchFlow, a Digital Twin for transient noise in Gravitational Wave Interferometers

GlitchFlow, a Digital Twin for transient noise in Gravitational Wave Interferometers Lorenzo Asprea 1 ,Elia Cellini 1,2 ,Federica Legger 1,∗ ,Alessio Romano 1 ,Francesco Sarandrea 1 , and Sara Vallero1 1Istituto Nazionale di Fisica Nucleare, Torino, Italy 2Università di Torino, Torino, Italy Abstract. Gravitational Waves (GW) were first predicted by Einstein in 1918, as a consequence of his theory of General Relativity published in 1915. The first direct GW detection was announced in 2016 by the LIGO and Virgo collaborations. Both experiments consist of a modified Michelson-Morley interferometer that can measure deformations of the interferometer arms of about 1/1,000 the width of a proton. The sensitivity of GW interferometers is limited by noise. Non-Gaussian transient noise artifacts, also known as glitches, are particularly challenging due to their similarity with astrophysical signals in the time and frequency domains. Noise reduction and subtraction is one of the most important and challenging activities in GW research. interTwin is a EU-funded project with the aim of building Digital Twins (DT) for various scientific use cases based on a co-designed blueprint architecture. Within interTwin, we are developing GlitchFlow, a pipeline for modeling and generating glitches for GW interferometers using deep generative algorithms. In this contribution, we present results of the glitch generation using several Neural Network (NN) architectures, and describe the implementation of the pipeline as execution of DAGs (Directed Acyclic Graph) with an Apache Airflow instance deployed on Kubernetes. We show how the most computing intensive tasks such as model training can be off-loaded to HPC resources, using interLink, a module developed within the interTwin framework. 1 Introduction Gravitational Waves (GWs) were long predicted by Albert Einstein’s theory of general relativity, but their first experimental evidence took almost a century [ 1 ]. GW observation on Earth requires highly sensitive interferometers that may detect waves generated by astrophysical processes such as the merging of black holes and neutron stars. The current generation detectors, Advanced Virgo [ 2 ] and its American counterpart Advanced LIGO (Laser Interferometer Gravitational-Wave Observatory) [ 3 ], use a laser-based modified Michelson-Morley interferometer with perpendicular arms. The laser beam is split into two beams, which travel down the arms and are reflected back by mirrors at the ends. The effective arm length is increased by additional mirrors inserted near the beam splitter to facilitate multiple reflections of the laser. When the beams recombine, a passing GW may be detected by observing changes in ∗e-mail: [email protected] EPJ Web of Conferences 337, 01203 (2025) https://doi.org/10.1051/epjconf/202533701203 CHEP 2024 © The Authors, published by EDP Sciences. This is an open access article distributed under the terms of the Creative Commons Attribution License 4.0 (https://creativecommons.org/licenses/by/4.0/). the interference pattern, allowing scientists to measure the wave’s amplitude and frequency. Advanced LIGO and Advanced Virgo have successfully observed multiple GW events since the first detection in 2015 [ 1 ]. The next-generation of ground-based interferometers, Einstein Telescope [ 4 ] (ET) in Europe and Cosmic Explorer [ 5 ] in the US, are being developed to further expand the frequency range that can be probed by the current detectors. Multi-Messenger Astronomy (MMA) [ 6 ] combines information from different types of astronomical signals, electromagnetic waves, GWs, neutrinos, and cosmic rays, to study astrophysical phenomena. GWs are fundamental to probe the dynamics of massive and compact objects, such as merging black holes or neutron stars. To detect potentially interesting astrophysical events, low-latency search pipelines analyze data from the interferometer network, which currently includes Virgo and LIGO. KAGRA [ 7 ], an underground interferometer in Japan, is expected to become operational in 2025 and join the network. The low-latency pipelines can identify a GW event candidate within approximately one minute of data acquisition [ 8 ]. Upon detection, an alert is sent to astronomers in order to search for the electromagnetic or neutrino counterparts. The sensitivity of GW interferometers is limited by various noise sources [ 9 ]. Seismic noise dominates at low frequencies, whereas thermal noise becomes significant at intermediate frequencies. High-frequency sensitivity is constrained by quantum noise, primarily due to photon shot noise [ 3 ]. Several techniques, such as the use of suspended mirrors, cryogenic cooling to reduce thermal motion, and quantum squeezing to mitigate shot noise, have been implemented to improve the instrument sensitivity. A significant experimental challenge is given by transient noise events, or glitches, since they may mimic the GW signal [ 10 ]. Transient noise may arise from a variety of sources, including environmental conditions, seismic activity, vibrations, electromagnetic interference, or instrumental artifacts within the interferometer. Each glitch has a unique signature, making it necessary to carefully characterize them, and to develop techniques to treat and distinguish them from measured GW events. An efficient and fast treatment of transient noise in the interferometers is therefore paramount for low-latency analyses that need to produce prompt and reliable alerts for other observatories. 2 The Virgo Digital Twin InterTwin [ 11 ] is an EU (European Union)-funded project aimed at co-designing and implementing a prototype of an interdisciplinary Digital Twin Engine (DTE). A Digital Twin (DT) is a virtual representation of a physical system or a process that simulates its real-world counterpart in real time. InterTwin aims to apply its DTE to several use cases from various fields, such as high-energy physics, radio astronomy, climate research, including GW astronomy with Virgo (and extension to ET). In this paper, we describe the development of the Virgo DT and how its components interact with the interTwin DTE developed modules. The Virgo DT aims to realistically simulate and detect transient noise (glitches) for low latency pipelines, i.e. within O (10 s ) after receiving the interferometer data, with the goal of providing a veto for events containing glitches and possibly de-noise the signal. This would allow alerts to be sent also for candidate GW events overlapping with a glitch. In the following we introduce a few concepts about GW experimental data analysis that are needed to understand the design of the Virgo DT. The output of ground-based GW detectors such as Virgo is a one-dimensional time series known as the strain amplitude, and referred to as h ( t ). The strain is a measurement of the differential length of the interferometer arms, and is therefore sensitive to a possible GW signal. Together with the strain, the so-called auxiliary channels are recorded. The auxiliary channels, in the order of several tens of thousands, contain time series measurements of the interferometer control systems and environmental EPJ Web of Conferences 337, 01203 (2025) https://doi.org/10.1051/epjconf/202533701203 CHEP 2024 2 sensors tracking conditions such as seismic activity, acoustic noise, and electromagnetic interference. The auxiliary channels and the strain are the main sources of information for noise characterization studies. The assumption behind the design of the Virgo DT is that both noise and GW events are recorded in the strain data, whereas there exist a subset of the auxiliary channels, called safe channels, which contain only noise (in addition to their specific monitoring measurement) but no GW candidate signal. Since the current ratio of observations of GW events in Virgo is a few tens of candidates per year, and each event may be observed for typically less than a minute, most of the strain data contain only noise. The aim of the Virgo DT is to find the mapping between a subset of auxiliary channels and the strain from data where no GW event has been recorded (only-noise case) using Generative Neural Network (GenNN) models. The output of the DT is the generated strain data in the noise-only case using the input from the selected auxiliary channels. The generated strain may be compared with the measured strain in order to either veto events containing glitches or to subtract the glitch contribution from a possible GW signal. Images are particularly suited as input of Machine Learning (ML) models such as Convolutional Neural Networks (CNN) and GenNNs. To leverage this, the one-dimensional time series - the strain data and the auxiliary channels that are inputs to the Virgo DT - are transformed into images using the Q-transform. The Q-transform [ 12 ] is a transformation often applied to the time series data in order to produce the signal energy distribution in the two-dimensional time-frequency plane. This time-frequency representation of the data allows for better visualization of the signal, that can be viewed as a two-dimensional image. The time-frequency maps help to highlight short-duration features in the data, and are often used in analysis tools for transient noise detection. A schematic representation of the Virgo DT is shown in Fig. 1. Figure 1. The Virgo DT schema. Data from the auxiliary channels are used to train a generative model, the GenNN, whose output is the generated strain data in the noise-only case (top). The Q-transforms of both generated and measured strain data are compared to show the transient noise glitch, visible in the generated image, and the superposition of the same glitch with a possible GW signal in the measured strain data (bottom). All images are drawn for visualization purposes and do not contain neither measured nor simulated Virgo data. To implement the Virgo DT we developed two main tools, Advanced Non-linear transientNoise Analyser of Laser Interferometer Sensor Array (ANNALISA) and GlitchFlow [ 13 ]. ANNALISA exploits the time-frequency domain representation of the data to evaluate correlations among the strain and auxiliary channels. GlitchFlow is a comprehensive framework containing all modules needed to train the GenNN model. EPJ Web of Conferences 337, 01203 (2025) https://doi.org/10.1051/epjconf/202533701203 CHEP 2024 3 2.1 ANNALISA ANNALISA leverages the time-frequency representation of the data to identify correlations aiming to reduce the number of auxiliary channels used as input for model training from several thousands to O (10 − 100). Specifically, it employs the Q-transform, a type of ShortTime Fourier Transform (STFT) with a varying window size. Instead of a fixed window, the Q-transform maintains a constant ratio (Q-value) between the center frequency and the frequency resolution. This results in high temporal resolution at high frequencies and high frequency resolution at low frequencies. This adaptive windowing makes the Q-transform particularly suited for capturing transient signals across a wide range of frequencies. The Q-transform generates a spectrogram where the x-axis represents time, the y-axis represents frequency, and the color intensity corresponds to the signal-to-noise ratio squared (SNR²), that can be understood as the normalized energy content of the signal at a particular time and frequency band [14]. ANNALISA computes correlations between Q-transforms of the input time series as the temporally coincident fraction of energy peaks, allowing for the detection of non-linear correlations among the input channels as peaks may appear at different frequencies in different channels. To make the correlation index more significant, only energy peaks above a certain threshold, which is set to 25 SNR², have been considered. We ran ANNALISA on the entire database of glitches classified as "scattered light" in Virgo data from the third observing run period (O3a), and selected eleven auxiliary channels with the highest correlations with the strain channel. Data from these channels are used as input for the GenNN model trained in the GlitchFlow pipeline. 2.2 GlitchFlow Virgo raw data are loaded as time series from local storage for preprocessing, retaining only the auxiliary channels selected by ANNALISA and the strain. The data is first whitened. Whitening refers to a preprocessing step that transforms a time-series signal so that its power spectrum becomes flat, making the data delta-correlated or Gaussian-like with uniform variance by removing all the correlation of the noise. As a second step, Q-transforms of the input data are computed. A Q-value of 12 has been chosen for the Q-transform calculation, as it has proven effective in capturing the transient nature of scattered light glitches in the frequency range of (8 − 500) Hz. The data is then divided into 64 × 64 pixel images, corresponding to 1 s× (8 − 500) Hz with logarithmic frequency spacing. The data is normalized with respect to the average maximum value of each channel. The preprocessed data is stored and used for training the GenNN. The images from the selected auxiliary channels are used as input to the GenNN, and the output image of the GenNN is compared to the target image from the measured strain. The current GenNN model architecture is a U-Net-inspired [ 15 ] encoder-decoder that also incorporates attention gates and residual blocks. The network is trained on 90% of the dataset, with the remaining 10% being used for validation. A loss combining Structural Similarity Index Measure [ 16 ] on generated image and L 1 distance between the target and generated images is employed for training with AdamW [ 17 ] optimizer and an initial learning rate of 10 −4 . The dataset is divided in batches of 200 samples, and the training is carried out on a Nvidia A100 GPU for 4 hours until the validation loss converges to a constant value. We evaluate the performance of the model for two purposes: issue a veto on events where a glitch is detected (vetoing), and perform denoising of the strain channel, defined as the pixel to pixel subtraction (cleaning) of the generated image from the target image. To perform vetoing, the model needs to reliably generate images where a glitch is detected if a glitch is present in the target image. A glitch is defined as any pixel cluster above a certain threshold SNR² value. For EPJ Web of Conferences 337, 01203 (2025) https://doi.org/10.1051/epjconf/202533701203 CHEP 2024 4 our validation procedure, we used a minimum cluster size of a single pixel. We measure the vetoing accuracy as the ratio of correctly predicted glitches over a validation set consisting of images with a 80% − 20% ratio of glitches - non glitches. The vetoing accuracy is above 90% for SNR² above 36. For effective denoising, the generated glitches need to have the correct position and shape in the time-frequency plane, in addition to the right intensity. We measure the denoising accuracy as the ratio of detected glitches in the validation set after cleaning. The results are shown in Fig. 2. Figure 2. Left: denoising accuracy as a function of the SNR² threshold used to define a glitch. Right: Examples of target (measured) strain, generated strain, and cleaned (after pixel to pixel subtraction) strain. The GlitchFlow pipeline is currently structured in a series of Jupyter notebooks designed to be integrated into the Virgo DT architecture, as explained in Section 3.1. 3 The Virgo DT architecture Building upon the interTwin DTE, we are implementing two subsystems for the Virgo DT: the Training subsystem and the Inference subsystem, respectively responsible for training and inference of the generative model as described in 2.2. These subsystems are operated by a DT Operator, which can be a person or an automated procedure that operates the DT during data-taking. The DT Operator controls the execution of the Training and Inference subsystems, and monitors the operations of the DT by checking relevant metrics on training convergence and inference accuracy. The trained GenNN models are saved and retrieved from an external model catalog, which store the history of the training runs, model parameters and hyper-parameters. 3.1 The DT Training Subsystem The main component of the DT Training Subsystem (TS) is the Data Store. The Data Store is used to store Virgo data in the form of time series, containing the strain channel and a subset of auxiliary channels selected by ANNALISA (see Section 2.1). In normal operating conditions the Data Store acts as a FIFO buffer, receiving an incoming stream of data from the interferometer, however in the current prototype implementation of the DT data are sent to the Data Store as files. The length of the buffer is currently under study, but we foresee to use about a one-month equivalent of data, corresponding to approximately 150 TB. When the buffer is full, the DT Operator triggers the training of the GenNN using data from the Data Store periodically (for example every month) or upon certain conditions (change of data taking conditions, new noise sources). EPJ Web of Conferences 337, 01203 (2025) https://doi.org/10.1051/epjconf/202533701203 CHEP 2024 5 The TS is composed by modules implemented as Directed Acyclic Graphs (DAGs), sets of instructions that are executed in predefined sequences without loops or cycles, and by events. Events are used to trigger DAG execution and to communicate with the TS external components (the DT Operator, the model catalog, the monitoring interface, and the storage volumes that hold the detector data). Each module contains a dedicated set of API (Application Programming Interface): • Data Store API. It allows the DT operator or other modules to interact with the Data Store. Upon receiving the “Start” event trigger, it sends a “Freeze” event to the Data Store, freezing the current data in the buffer, i.e. no new data will be written in the Data Store, and no data will be deleted. When the Data Store enters the “Frozen” condition, the Data Store API sends a “Frozen” event to the Preprocessing API. • Preprocessing API. The Preprocessing module spawns upon reception of the “Frozen” event from the Data Store API. The Preprocessing module reads data from the Data Store and builds time series of the appropriate length for training (using crop or join methods). It applies standard-preprocessing (de-noising) and Q-transform to produce time-frequency spectrograms, and then turns them into 2D images. Preprocessed data are stored in a temporary object store. • Training API. This is a Python-based module that trains the GenNN model on preprocessed data, and stores the trained model in an external Model Catalog. Its execution is triggered by receiving the event "Train" by the Preprocessing API. The architecture and data flow of the Virgo DT TS is shown in Fig. 3. Figure 3. The Virgo DT Training Subsystem. The internal modules and the events triggering DAG execution as described in Section 3.1 are included in the orange box. A legend is provided at the bottom. External components such as the DT Operator, the Model Catalog, and monitoring are also shown. 3.2 The DT Inference Subsystem The Virgo DT Inference subsystem has a similar architecture to the Training subsystem described in Section 3.1, but it processes a stream of incoming interferometer data, rather than using data from the Data Store buffer. The incoming data contain time series of the Virgo strain channel and the relevant auxiliary channels. The data is preprocessed by the Preprocessing module to produce images in the same format as those used by the Training module. Inference is applied using the latest GenNN model retrieved from the Model Catalog. The aim of this component is to identify glitches and issue a decision about further processing (veto or denoising). EPJ Web of Conferences 337, 01203 (2025) https://doi.org/10.1051/epjconf/202533701203 CHEP 2024 6 4 Current status and future work InterTwin is a 36-month long project that started on September 1st, 2022. At the time of this writing, first versions of the main components of the Virgo DT, ANNALISA and GlitchFlow, have been released in the public interTwin repository [ 18 ]. A first release of the Virgo DT Training subsystem is also available [ 19 ]. It consists in a working prototype with an implementation of the Data Store and the Data Store APIs. The DAGs are written in Python and their execution is orchestrated by Apache Airflow using HTTP sensors. Apache Airflow is deployed on a Kubernetes cluster. The DAG execution is monitored using the Airflow provided dashboard, and logs are collected with Logstash and stored in ElasticSearch. Currently the Preprocessing and Training APIs consist of dummy modules, and only the API interfaces exist. At a later stage the modules will be filled with full functionalities from the GlitchFlow framework. Future work will also include the implementation of the DT Virgo Inference subsystem, based on the modules already developed for the Training subsystem. The interTwin-developed component interLink [ 20 ] will be used to schedule the execution of time-consuming pods, such as those spawned by the Preprocessing module or the Training module, on external resources such as HPCs or clouds. A pod is the basic unit of Kubernetes capable of executing a workload, and in its simplest form it can be though of as a container. InterLink provides an abstraction layer for the execution of a Kubernetes pod on any remote resource capable of managing a container execution life cycle. The Virgo interferometer data will be made available on remote computing resources through Rucio [ 21 ], the framework chosen by interTwin for data management and distribution. Further exploitation of the interTwin DTE is being carried out by integrating GlitchFlow with itwinai [ 22 ], a toolkit used to compose modular and reusable ML workflows. This will allow us to scale out embarrassingly parallel tasks such as hyper-parameter tuning on HPC resources. 5 Summary We presented an overview of the Virgo DT in the context of the EU project interTwin, which ends on Aug. 31st, 2025. We described the novel implementation of the Training and Inference subsystems for the Virgo DT, and the development of two new tools, ANNALISA and GlitchFlow. Most of the components described in this work are already publicly available. Further work is needed to finalise the Inference subsystem, complete the integration with the interTwin DTE, and fully exploit the real-time capabilities of the Virgo DT. Acknowledgments L.A., F.S. and S.V. acknowledge support from the interTwin project funded by the European Union Grant Agreement Number 101058386. E.C. acknowledges support from the SFT Scientific Initiative of INFN and the Simons Foundation grant 994300 (Simons Collaboration on Confinement and QCD Strings). References [1] The LIGO Collaboration and Virgo Collaboration, Observation of Gravitational Waves from a Binary Black Hole Merger. Phys. Rev. Lett. 116, 061102 (2016), https://doi.org/10. 1103/PhysRevLett.116.061102 [2] F. Acernese et al., Advanced Virgo: a second-generation interferometric gravitational wave detector. Class. Quantum Grav. 32 024001 (2015), https://doi.org/10.1088/0264-9381/ 32/2/024001 [3] The LIGO Scientific Collaboration et al., Advanced LIGO. Class. Quantum Grav. 32 074001 (2015), https://doi.org/10.1088/0264-9381/32/7/074001 EPJ Web of Conferences 337, 01203 (2025) https://doi.org/10.1051/epjconf/202533701203 CHEP 2024 7 [4] M. Punturo et al., The Einstein Telescope: a third-generation gravitational wave observatory. Class. Quantum Grav. 27 194002 (2010), https://doi.org/10.1088/0264-9381/27/19/ 194002 [5] M. Evans et al., A Horizon Study for Cosmic Explorer: Science, Observatories, and Community. arXiv:2109.09882 [astro-ph.IM] (2021), https://doi.org/10.48550/arXiv.2109. 09882 [6] M. Branchesi, Multi-messenger astronomy: gravitational waves, neutrinos, photons, and cosmic rays. J. Phys.: Conf. Ser. 718 022004 (2016), https://doi.org/10.1088/1742-6596/ 718/2/022004 [7] T.Akutsu et al., Overview of KAGRA: Detector design and construction history. Prog. Theor. Exp. Phys., 05A101 (2021), https://doi.org/10.1093/ptep/ptaa125 [8] B. P. Abbott et al., Low-latency Gravitational-wave Alerts for Multimessenger Astronomy during the Second Advanced LIGO and Virgo Observing Run. Astrophys. J. 875/2 161 (2019), https://doi.org/10.3847/1538-4357/ab0e8f [9] B. P. Abbott et al., A guide to LIGO–Virgo detector noise and extraction of transient gravitational-wave signals. Class. Quantum Grav. 37 055002 (2020) https://doi.org/10.1088/ 1361-6382/ab685e [10] B. P. Abbott et al., Characterization of transient noise in Advanced LIGO relevant to gravitational wave signal GW150914. Class. Quantum Grav. 33 134001 (2016), https: //doi.org/10.1088/0264-9381/33/13/134001 [11] The interTwin website,https://www.intertwin.eu/, accessed: 2024-01-18 [12] S. Chatterji, et al., Multiresolution techniques for the detection of gravitational-wave bursts. Class. Quant. Grav. 21 S1809-S1818 (2004), https://doi.org/10.1088/0264-9381/21/ 20/024 [13] Y. Pidopryhora, et al., interTwin D4.4 First version of the DTs capabilities for High Energy Physics, Radio astronomy and Gravitational-wave Astrophysics. Zenodo (2024), https://doi.org/10.5281/zenodo.11403139 [14] F. Acernese et al., Virgo detector characterization and data quality: tools. Class. Quant. Grav. 40,18, 185005 (2023), https://doi.org/10.1088/1361-6382/acdf36 [15] O. Ronneberger et al., U-Net: Convolutional Networks for Biomedical Image Segmentation. Lecture Notes in Computer Science 9351:234-241, (2015) http://dx.doi.org/10.1007/ 978-3-319-24574-4_28 [16] Z. Wang, et al., Image quality assessment: from error visibility to structural similarity. IEEE Trans. Image Process., 13(4) 600-612 (2004) https://doi.org/10.1109/TIP.2003. 819861 [17] I. Loshchilov and F. Hutter., Decoupled Weight Decay Regularization. International Conference on Learning Representations (ICLR 2019), https://doi.org/10.48550/arXiv. 1711.05101 [18] GlitchFlow,https://github.com/interTwin-eu/DT-Virgo-notebooks, accessed: 2024-0118 [19] The DT Virgo DAGs,https://github.com/interTwin-eu/DT-Virgo-dags/tree/main/Release, accessed: 2024-01-18 [20] interLink,https://intertwin-eu.github.io/interLink/, accessed: 2024-01-18 [21] M. Barisits, T. Beermann, F. Berghaus, et al., Rucio: Scientific Data Management. Comput. Softw. Big Sci. 3: 11 (2019), https://doi.org/10.1007/s41781-019-0026-3 [22] itwinai,https://github.com/interTwin-eu/itwinai, accessed: 2024-01-18 EPJ Web of Conferences 337, 01203 (2025) https://doi.org/10.1051/epjconf/202533701203 CHEP 2024 8