Focus on:  Hide Contributions
Back to Conference View
Login

AI for Gravitational Waves Workshop @ CERN

→ Europe/Zurich
40/S2-A01 - Salle Anderson (CERN)

40/S2-A01 - Salle Anderson

CERN

95
Show room on map
Eric Anton Moreno (Massachusetts Institute of Technology (US)), Katya Govorkova (Massachusetts Inst. of Technology (US)), Maurizio Pierini (CERN)
Description

AI for Gravitational Waves Workshop

This workshop is on how AI and machine learning are reshaping gravitational wave science across four themes: Data Analysis (low-latency detection, parameter inference, transient searches, interpretability), GW Simulation (fast waveforms, surrogate modeling, differentiable simulation, uncertainty quantification), Real-Time Data Processing (streaming ML, scalable inference, GPU/FPGA acceleration, model compression), and Detector Operations (monitoring, noise hunting, data quality, predictive maintenance).

The programme features keynote talks from the LVK, ET, LISA, MMA, and PTA, bringing together ML researchers, GW astronomers, instrumentalists, and computing experts to surface shared challenges and collaboration opportunities ahead of the next observing runs and future missions.

Invited speakers

  • Viola Sordini (IPN, Lyon)
  • Michele Valisneri (ETH, Zurich)
  • Paolo Pani (Sapienza University of Rome)
  • Stas Babak (CNRS, Paris)
  • Daniel Muthukrishna (MIT, US)
  • Max Dax (ELLIS Institute and MPI, Tübingen)
  • Deep Chatterjee (MIT)
  • Michael Coughlin (University of Minnesota)
  • Huw Haigh (Austrian Academy of Sciences)
  • James Alvey (University of Cambridge)
  • Melissa Lopez (Nikhef and Utrecht University)
  • Uddipta Bhardwaj (ETH)
  • Tomislav Andric (Gran Sasso Science Institute)
  • Nikhil Mukund (MIT)
  • Christina Reissel (MIT)
  • Jonathan Klimesch (University of Tübingen)
  • Siddharth Soni (University of California)

Scientific committee

  • Elena Cuoco (University of Bologna)

  • Valerie Domcke (CERN)

  • Jan Harms (Gran Sasso Science Institute)

  • Gianluca Inguglia (Austrian Academy of Sciences)

  • Erik Katsavounidis (MIT)

  • Samaya Nissanke (DESY, German Centre for Astrophysics DZA)

Organising committee

  • Katya Govorkova (MIT)

  • Eric Moreno (MIT)

  • Maurizio Pierini (CERN)

 

AI4GW_at_CERN.pdf
AI for Gravitational Waves Workshop Booking form.docx
AI for Gravitational Waves Workshop Booking form.pdf
CERN-logo_outline[1].jpg
Contact
Participants
173 View full list
AI for Gravitational Waves Workshop @ CERN
Zoom Meeting ID
62734231983
Host
Alex Lasa Lamarca
Useful links
Join via phone
Zoom URL
  • Tue, May 5
  • Wed, May 6
    • 9:00 AM → 10:20 AM
      AI for data analysis: Contributed Talks
      Convener: gianluca inguglia
      • 9:00 AM
        Real-Time Gravitational-Wave Inference with Probabilistic Machine Learning 20m

        Gravitational-wave (GW) astronomy promises groundbreaking discoveries in the coming decades, but its progress is bottlenecked by the computational challenges of large-scale and real-time data analysis. I will present DINGO, a machine learning approach for fast and accurate GW inference that addresses these challenges. DINGO trains generative neural networks to directly estimate probability distributions over GW source parameters. I will explain the core ideas behind DINGO and highlight several machine learning techniques that we developed to adapt modern simulation-based inference to the challenging field of GW data analysis.

        Speaker: Maximilian Dax
        Recording
        slides_dax.pdf
      • 9:20 AM
        Using AI for GW detection, Classification and Parameter Estimation 20m

        We present a study of deep convolutional autoencoders applied to anomaly detection of GW signals. This initial work focuses on short-duration signals (< 2s), corresponding to mergers that involve, or form, intermediate mass black holes. These burst-like signals are notably difficult to disentangle from both background noise and glitches that may occur during data taking. We utilise the simulated noise and merger catalogue provided as part of the Einstein Telescope Mock Data Challenge. Weak supervision is employed during training, whereby the model is directly optimised to separate 2D spectrograms containing IMBH merger signals (injected into ET noise) from those containing only noise.

        The model shows excellent results in recovering the targeted IMBH merger signals, and a strong ability to generalise to masses below those seen during training. Current work focuses on the inclusion of glitches in training, with more complex network architecture being tested to provide 3-way noise, glitch and signal classification. Furthermore, the inclusion of networks purposed towards parameter estimation is under investigation, with the aim to thereby develop a classification and parameter estimation pipeline that is able to handle the high rate and diversity of signals we can expect in the Einstein Telescope era.

        Speaker: Huw Haigh (Austrian Academy of Sciences (AT))
        CNN based anomaly detection - CERN 2026.pdf
        CNN based anomaly detection - CERN 2026.pptx
        Recording
      • 9:40 AM
        Rapid Detection and Inference of Extreme-Mass-Ratio Inspirals in LISA: Divide and Conquer 20m

        Extreme-mass-ratio inspirals (EMRIs) are key gravitational-wave sources for the Laser Interferometer Space Antenna (LISA), but their detection and parameter inference are computationally challenging due to the extreme concentration of posterior distributions within vast prior volumes. In this work, we introduce a novel divide-and-conquer strategy that reformulates global inference as a hierarchical identification problem. Our approach iteratively localizes the posterior mode through a coarse-to-fine procedure based on ordinal classification, progressively restricting the parameter space while preserving the true signal region. A transformer-based neural network is trained at each stage to identify the most probable parameter subregions, enabling exponential reduction of the search volume with only a few refinement steps. Once the parameter space is reduced to the Fisher-information scale, standard local sampling methods efficiently recover the full joint posterior. We demonstrate that this method achieves rapid and accurate intrinsic parameter estimation for EMRIs in simulated LISA data, dramatically reducing computational costs compared to traditional global sampling techniques. This framework provides a scalable and efficient pathway for real-time EMRI detection and inference in the LISA era.

        Speaker: Rahul Srinivasan
        gg
        gg
        Recording
      • 10:00 AM
        The power of Normalizing Flows for Bayesian inference 20m

        Pulsar Timing Array data analysis faces severe computational challenges as parameter spaces scale with the number of pulsars. I present two Normalizing Flows (NFs) based strategies to accelerate and improve Bayesian inference for stochastic gravitational wave background (SGWB). First, integrating NFs into the importance nested sampling framework i-nessai yields speedups of one to three orders of magnitude over standard methods, with robust posteriors and reliable evidence estimates. Second, a dual NFs architecture implementing parameter decorrelation via orthogonal projection disentangles pulsar noise from hyperprior parameters in hierarchical Bayesian modeling, enhancing noise constraining power even in the presence of a SGWB signal.

        Speaker: Dr Eleonora Villa (INAF - IASF Mi)
        E.Villa-CERN.pdf
        Recording
    • 10:20 AM → 10:50 AM
      Coffee Break 30m
    • 10:50 AM → 12:10 PM
      AI for data analysis: Contributed Talks
      Convener: gianluca inguglia
      • 10:50 AM
        Tackling the LISA Global Fit: Scalable Simulation-Based Inference and the Road Ahead 20m

        The Laser Interferometer Space Antenna (LISA) will deliver an unprecedented view of the gravitational-wave universe, but unlocking its scientific potential hinges on a monumental data science challenge: the "global fit". Extracting thousands of overlapping, time-varying signals from complex instrumental noise is a high-dimensional inference problem that pushes the limits of traditional stochastic sampling methods. In this talk, I will explore how Simulation-Based Inference (SBI) is emerging as a promising framework to break this computational bottleneck. I will overview the current landscape of SBI applied to LISA data, highlighting how it has been used across the various source classes. Finally, I will outline what I see as the significant hurdles that remain as we scale these algorithms toward a full analysis pipeline.

        Speaker: James Alvey (University of Cambridge)
        2026-05-06 AI4GW.pdf
        Recording
      • 11:10 AM
        ★ Simulation-Based Homotopy: Stress-Testing Gravitational-Wave Posteriors ★ 20m

        We present a framework for probing the full geometry of Bayesian posteriors in inverse problems through a noise-conditioned homotopy. By embedding the likelihood in a one-parameter family controlled by a noise-scaling parameter, we construct a continuous deformation from an almost deterministic posterior concentrated at the true parameters to the full noisy posterior.
        Traversing this path reveals how posterior structure evolves with measurement quality: when multi-modality emerges, where Gaussian approximations break down, and how parameter degeneracies develop. We argue this constitutes a more integrated alternative to Fisher-information analyses, which becomes beneficial especially in multimodal geometries.

        Additionally, deviations from smooth homotopy behaviour provide direct diagnostics of inference pipelines, allowing identification of spurious correlations, mode-collapse artefacts, and approximation breakdowns. We discuss the framework as a general validation and benchmarking tool for simulation-based inference methods.

        Speaker: Giada Chiara Badaracco (ETH Zurich (CH))
        posterior_homotopy_talk.pdf
        posterior_homotopy_talk.pptx
        Recording
      • 11:30 AM
        ★ Overlapping signals in 3G detectors: an approach based on Transformers ★ 20m

        Third-generation ground-based gravitational wave detectors such as the Einstein Telescope are expected to significantly advance our understanding of compact binary coalescences. One of the most critical challenges in data analysis for the Einstein Telescope is that of overlapping signals. With a tenfold improvement in sensitivity, the Einstein Telescope will be able to detect binary black hole and binary neutron star coalescences with expected rates of up to ~10⁵ events per year. Moreover, the extended range toward lower frequencies will allow the detector to observe these signals for longer durations compared to current-generation detectors. While this creates the opportunity to deepen our knowledge of these sources, detectable signals will inevitably overlap. This poses a severe challenge to parameter estimation analysis pipelines. We need a faster, unbiased parameter estimation strategy.

        In this talk, we will describe a promising solution to address this challenge: a deep learning approach that combines the power of two state-of-the-art machine learning architectures, Transformers and Normalizing Flows. In particular, we present the first application of a Transformer encoder for gravitational wave data analysis. This architecture is capable of capturing complex, varying-range dependencies, and we use it to extract the information in the data. We then employ Normalizing Flows to estimate the high-dimensional posterior distributions of the overlapped signals.
        We will present the results from training this network architecture, demonstrating its effectiveness in handling three overlapping signals simultaneously, and discussing how this deep learning method represents a promising solution to the problem, along with its potential extensions and improvements.

        Speaker: Lucia Papalini (University of Pisa & INFN-Pisa)
        ai4gw-cern.pdf
        ai4gw-cern.pptx
        Recording
      • 11:50 AM
        Accelerating LISA inference with Gaussian processes 20m

        Source inference for deterministic gravitational waves is a computationally demanding task in LISA. In a novel approach, we investigate the capability of Gaussian processes to learn the posterior surface in order to reconstruct individual signal posteriors. We use GPry, which automates this reconstruction through active learning, using a very small number of likelihood evaluations, without the need for pretraining. We benchmark GPry against the cutting-edge nested sampler nessai, by injecting individually three signals on LISA noisy data simulated with Balrog: a white dwarf binary (DWD), a stellar-mass black hole binary (stBHB), and a supermassive black hole binary (SMBHB). We find that GPry needs $\mathcal{O}(10^{-2})$ fewer likelihood evaluations to achieve an inference accuracy comparable to nessai, with Jensen-Shannon divergence $D_{JS}≲0.01$ for the DWD, and $D_{JS}≲0.05$ for the SMBHB. Lower accuracy is found for the less Gaussian posterior of the stBHB: $D_{JS}≲0.2$. Despite the overhead costs of GPry, we obtain a speedup of $\mathcal{O}(10^2)$ for the slowest cases of stBHB and SMBHB. In conclusion, active-learning Gaussian process frameworks show great potential for rapid LISA parameter inference, especially for costly likelihoods, enabling suppression of computational costs without the trade-off of approximations in the calculations or expensive model pre-training.

        Speaker: JESUS TORRADO CACHO (Instituto de Estructura de la Materia (IEM-CSIC))
        Accelerating LISA inference with Gaussian Processes – Jesús Torrado.pdf
    • 12:10 PM → 1:20 PM
      Lunch 1h 10m
    • 1:20 PM → 3:00 PM
      AI for GW Simulation
      Convener: Erik Katsavounidis (MIT)
      • 1:20 PM
        Representation-Aware Self Supervised Learning for Gravitational Wave Science: Introducing GWCLIP 20m

        The next generation of gravitational wave detectors will produce data at a scale and complexity that renders traditional matched-filtering and parameter estimation pipelines insufficient as standalone tools. We argue that the field must shift toward representation-aware learning — building models that acquire rich, physics-informed embeddings of gravitational wave signals rather than optimising narrowly for a single downstream task. Inspired by contrastive multimodal frameworks such as CLIP, we introduce GWCLIP: a model trained to align gravitational wave strain data with their corresponding physical descriptions — waveform families, source parameters, and detector context — through contrastive objectives across a large and diverse signal corpus.

        This shared embedding space enables zero/few-shot classification, rapid similarity search across catalogs, and robust transfer to downstream tasks including glitch rejection, denoising and parameter inference with minimal labeled data. We demonstrate that GWCLIP embeddings capture physically meaningful structure, clustering signals by source parameters in a geometry that generalises across detector configurations.

        Speaker: Uddipta Bhardwaj (ETH Zurich)
        GWCLIP.pdf
        Recording
      • 1:40 PM
        Fast Simulation of Gravitational-Wave Signals and Glitch Populations for Data Analysis 20m

        Accurate signal models and the disruptive presence of transient noise artifacts (glitches) impose key limitations on the sensitivity of gravitational-wave detectors. Efficient simulation of both is essential for data analysis. Fast generation of gravitational-wave signals is required for detection and parameter estimation, where waveforms are evaluated repeatedly within inference pipelines. At the same time, realistic modeling of glitches is needed to represent unknown noise populations and to stress test search and inference methods.
        This talk compares state-of-the-art approaches for fast simulation of both gravitational-wave signals and glitches. We focus on generative deep learning methods that learn data distributions and enable rapid generation in the time domain. These methods speed up simulations compared to traditional approaches while preserving key features of signals and noise, enabling more efficient and realistic data analysis.

        Speaker: Melissa Lopez (Utrecht University)
        Recording
      • 2:00 PM
        ★ What Lies Beneath the Noise: Inferring Galactic Binary Populations in LISA with SBI ★ 20m

        The LISA space mission, set to launch in the mid 2030s, will open a new window on the “gravitational wave universe”. Thanks to its exceptional sensitivity in the low frequency band ~10⁻⁴–10⁻¹ Hz, it will observe a variety of sources all at the same time: from massive black hole binaries to extreme mass ratio inspirals and Galactic compact binaries. Among these, double white dwarf binaries, emitting nearly monochromatic signals, are expected to dominate the mHz band. Of the ~10⁷ binaries in the Milky Way, only a small fraction—ranging from tens to thousands—will be individually resolvable by LISA, while the unresolved population will produce a stochastic foreground (“confusion noise”) that must be accounted for in the analysis of other sources.

        Different astrophysical formation and evolution channels are expected to leave distinct imprints on this confusion noise. However, extracting population-level information from this signal remains a challenging and computationally expensive task within the standard Global Fit framework.

        In this work, we present a machine learning approach based on Simulation-Based Inference to directly link the confusion noise to the underlying astrophysical population. We develop fast simulators to efficiently generate synthetic catalog realizations, and train a Neural Posterior Estimator to learn the mapping from the observed foreground to the parameters describing the binary population. We show that this approach can efficiently recover key population properties, bypassing some of the limitations of traditional inference pipelines.
        Our results highlight the potential of modern AI-driven inference methods for gravitational-wave data analysis, and represent a step towards integrating Simulation-Based Inference within the LISA Global Fit framework.

        This work is presented in detail in arXiv:2602.18560

        Speaker: Federico De Santi (University of Milano-Bicocca)
        AI4GW_CERN_2026_slides.pdf
        Recording
      • 2:20 PM
        Comparing next-generation detector configurations for high-redshift gravitational wave sources with neural posterior estimation 20m

        The coming decade will be crucial for determining the final design and configuration of a global network of next-generation (XG) gravitational-wave (GW) detectors, including the Einstein Telescope (ET) and Cosmic Explorer (CE). In this study and for the first time, we assess the performance of various network configurations using neural posterior estimation (NPE) implemented in Dingo-IS-a method based on normalizing flows and importance sampling that enables fast and accurate inference. We focus on a specific science case involving short-duration, massive and high-redshift binary black hole (BBH) mergers with detector-frame chirp masses ($\mathcal{M}_\mathrm{d}$) > 100 M$_\odot$. These systems encompass early-Universe stellar and primordial black holes, as well as intermediate-mass black-hole binaries, for which XG observatories are expected to deliver major discoveries. Validation against standard Bayesian inference demonstrates that NPE robustly reproduces complex and disconnected posterior structures across all network configurations. For a network of two misaligned L-shaped ET detectors (2L MisA), the posterior distributions on luminosity distance can become multimodal and degenerate with the sky position, leading to less precise distance estimates compared to the triangular ET configuration. However, the number of sky-location multimodalities is substantially lower than the eight expected with the triangular ET, resulting in improved sky and volume localization. Adding CE to the network further reduces sky-position degeneracies, and the better performance of the 2L MisA configuration over the triangle remains evident.

        Speaker: Filippo Santoliquido (Gran Sasso Science Institute)
        2Ls_high_redshift.pdf
        Recording
      • 2:40 PM
        Autoencoder based surrogate model for gravitational waves from BBH mergers 20m

        The generation of accurate waveforms from binary black hole (BBH) mergers is a major effort in Gravitational-Wave Astronomy. In recent years, machine-learning-based surrogate models for BBH waveforms have been proposed. Those offer the potential to dramatically accelerate waveform generation while maintaining accuracy competitive with that of traditional waveform approximants. In this work, we investigate the viability of autoencoders as generative models for gravitational-wave signals from quasi-circular BBH mergers. We introduce AESur3dq8, a novel surrogate waveform model based on autoencoders that enables the rapid and accurate construction of large template banks, producing millions of waveforms in under a second using modest computational resources. The model is trained on the numerical-relativity-informed surrogate NRHybSur3dq8 and subsequently fine-tuned using the SXS catalog of BBH simulations. We demonstrate that waveforms generated by AESur3dq8 achieve mismatches of order $10^{-4}$ with respect to Numerical Relativity waveforms, and that parameter estimation performed with these templates yields results fully consistent with those reported by the LIGO-Virgo-KAGRA Collaboration for observed gravitational-wave events.

        Speaker: Anastasios Theodoropoulos
        Autoencoder_Theodoropoulos_CERN_final.pdf
        Recording
    • 3:00 PM → 3:10 PM
      Group photo in the coffee area
    • 3:10 PM → 4:20 PM
      Posters: Posters and coffee
      • 3:10 PM
        Accurate and efficient simulation-based inference for massive black-hole binaries with LISA 1h 10m

        We develop an accurate simulation-based inference framework for high-mass ($\gtrsim\!10^7 \rm{M_\odot}$) black-hole binaries observable by LISA. The method is implemented within the DINGO gravitational-wave parameter-estimation code, extending its application from ground-based detectors to the LISA band. We train a normalizing-flow model using aligned-spin higher-mode waveform models and a low-frequency approximation of the detector response. After sampling, we importance-sample to the true posterior. We validate performance on simulated signals spanning the signal-to-noise regimes relevant for LISA observations and benchmark our new DINGO implementation against standard methods. We report robust agreement in the inferred posterior distributions up to signal-to-noise ratios of ∼ 500. At higher signal-to-noise ratios of ∼ 1000, we observe a reduction in sampling efficiency, while still yielding unbiased and tightly localized posteriors that can be used as a starting point for follow-up with traditional methods. The trained flow can generate 20 thousand posterior samples in less than a minute, establishing DINGO as a promising neural inference framework for rapid full-parameter estimation of massive black-hole binaries in the LISA band. The likelihood-free nature of this approach allows for straightforward generalizations, including a time-dependent detector response, non-stationary noise artifacts such as gaps and glitches, and low-latency parameter estimations.

        Speaker: Alice Spadaro (L2IT)
        PosterDINGO.pdf
      • 3:10 PM
        Calibrated GW Bayes Factors in real noise 1h 10m

        Bayesian analyses have long been at the forefront of GW data analyses, mostly using nested sampling. With even state-of-the-art simulation-based-inference methods working in a Bayesian regime, they are here to stay. This has over time lead to discussions about new GW discoveries, be it general-relativity precession, eccentricity, or population-level model comparisons, to be held almost exclusively through the lens of calculating and comparing Bayes Factors. However, these factors are only as valid as the assumptions that allow us to use Bayesian methods in the first place. Currently, none of those assumptions can deal with real detector noise, be it large bias-inducing glitches or just low-SNR non-Gaussianities. This is most evidently shown in the fact that most LIGO-Virgo-KAGRA events with claims of precession, eccentricity, or even lensing, are known to also contain glitches. This talk presents results from a new method using so-called "evidence networks", which statistically determine Bayes Factors based on nothing but the data distribution and can be amortized over real noise. We show that we can re-create nested-sampling results in idealized-noise regimes, and how quickly answers start to diverge in real noise. We plan to use these methods to cross check previous claims and to search for yet-undiscovered effects in LIGO-Virgo-KAGRA data.

        Speaker: Alexandre Göttel (School of mathematical sciences, University of Nottingham)
      • 3:10 PM
        Deploying GeoAI for Siting Cosmic Explorer 1h 10m

        Recent work on identifying suitable sites for the next-generation gravitational wave observatory, Cosmic Explorer (CE), includes site visits, efforts to build relationships with relevant communities surrounding promising sites, characterizing costs of construction, and finding suitable sites based on scientific thresholds. The common theme amongst these complementary efforts is Geographic Information Systems (GIS). GIS, like many other forms of data science, has a blooming community of researchers integrating their domain expertise with Artificial Intelligence (AI) to find spatial patterns in noisy data with non-linear relationships. In GIS, deploying AI into spatial analysis is referred to as GeoAI. With promising sites for Cosmic Explorer identified at the national scale, local analysis requires the use of GeoAI [DS1.1]to forecast future sources of anthropogenic noise as well as environmental hazards to Cosmic Explorer at potentially promising sites. This presentation explores how GeoAI will be used in local-scale and site level analysis to predict future land use (a principal concern for anthropogenic noise) and flood hazard simulation mapping in often remote, sparsely mapped portions of the contiguous United States. GeoAI will be key to bring CE to realization as local-scale and site level analysis will require spatial analysis that utilizes neural networks and deep learning.

        Speaker: Warren Bristol (University of Arizona)
      • 3:10 PM
        Early Detection of Sudden Beam Loss at SuperKEKB using Time-Series Anomaly Detection 1h 10m

        Beam loss events at SuperKEKB represent a major operational challenge, threatening both the Belle II detector and accelerator components while significantly impacting data-taking rates, with diamond doses reaching several thousand mrad per event.
        The current abort system relies on loss monitors positioned outside the beam pipe and at the interaction point. However, for Sudden Beam Loss (SBL) events, where significant current can be lost within a single turn, faster abort triggers are required.
        We show that bunch oscillation recorders (BOR), which provide turn-by-turn position data at ~200 MHz, contain precursor signals that can be utilized for early warning.

        We develop and compare three classes of anomaly detection algorithms on BOR data: sliding-window variance, Kalman filters, and a multivariate DeepLSTM forecasting model. We demonstrate that all three methods can successfully identify oscillation precursors ahead of conventional loss monitors, with the multivariate DeepLSTM offering the best performance. These findings establish the feasibility of BOR-based early warning triggers for beam loss mitigation at SuperKEKB.

        Speaker: Philipp Horak (University of Victoria)
      • 3:10 PM
        Enhancing DeepClean: Towards Scalable and Adaptive Gravitational-Wave Denoising 1h 10m

        This project builds on DeepClean, a machine learning framework for subtracting instrumental and environmental noise from gravitational-wave detector data using auxiliary “witness” channels. While effective, DeepClean’s convolutional architecture limits its ability to scale across varying channel configurations and frequency bands.
        We propose a hybrid CNN–transformer architecture that maintains permutation invariance across channels. The model combines a per-channel CNN for independent feature extraction with a transformer that treats time steps as tokens and models cross-channel interactions via self-attention, while preserving channel-specific representations. This design aims to reduce retraining requirements under changing detector configurations, such as varying channel sets or frequency bands.
        Preliminary results show performance comparable to DeepClean on standard denoising benchmarks, indicating that this added flexibility does not compromise accuracy. Ongoing work explores improved performance across broader frequency bands and increased robustness to imperfect or noisy channel selection. Ultimately, this project aims to move toward a more automated version of DeepClean, incorporating learning-based channel selection tailored to the denoising task.

        Speaker: Sharvari Tatachar (Georgia Institute of Technology)
        Recording
        STatacharPoster.pdf
      • 3:10 PM
        From Fast Radio Bursts to Gravitational Waves: Adapting Multi-Modal CNNs for Robust, Low-Latency Transient Detection 1h 10m

        As next-generation observatories push the boundaries of sensitivity and data throughput, the challenge of isolating faint, unmodeled astrophysical transients from non-Gaussian terrestrial noise has become a universal bottleneck. Both radio astronomy and gravitational-wave (GW) physics face highly analogous data environments: continuous, high-rate time-series streams plagued by instrumental glitches and interference that easily mimic true signals.

        This presentation outlines a mature, edge-deployable machine learning architecture originally developed for the real-time detection of Fast Radio Bursts (FRBs) at remote outrigger stations, and maps its direct translatability to GW low-latency pipelines. We detail a Multi-Input Convolutional Neural Network (CNN) designed to overcome the limitations of single-domain representation. Rather than relying solely on a 2D spectrogram, our architecture utilizes a branched topology to concurrently extract features from three distinct physical representations: the 1D integrated time profile, the 1D bandpass spectrum, and the 2D temporal-spectral waterfall plot.

        By forcing the network's dense decision layers to cross-reference multi-modal physics, the pipeline achieves exceptional robustness, drastically suppressing false-positive rates caused by localized terrestrial interference (which typically mimics a transient in only one domain). Furthermore, we discuss the coupling of this multi-branch CNN with Just-In-Time (JIT) compiled streaming buffers, ensuring the inference latency remains well within the strict bounds required for automated, multi-messenger follow-up prioritization. By bridging the data analysis methodologies of radio and GW astronomy, we present a scalable framework for unmodeled search representation learning, directly applicable to the noise-hunting and transient detection challenges of the LIGO-Virgo-KAGRA network and the upcoming Einstein Telescope.

        Speaker: Rizchel Masong
      • 3:10 PM
        GravCloud: A stacked LSTM-Pointcloud Framework for LIGO Glitch Classification 1h 10m

        Accurate classification of instrumental and environmental noise glitches is essential for gravitational-wave detector characterisation and data quality assurance in LIGO. We present GravCloud, a hybrid deep-learning framework that interprets one-dimensional transient noise burst glitches transformed into a multi-dimensional pointcloud representation, allowing noise transients to be analysed in a richer (geometric) space. The architecture combines Long Short-Term Memory (LSTM) networks for temporal feature extraction with Dynamic Graph Convolutional Neural Networks (DGCNN) for learning spatial relationships of the transient noise parameters represented in a higher dimensional plane. This approach is intended to leverage existing multi-dimensional structural learning tools to capture complex glitch characteristics more compute effectively than conventional one-dimensional or even image-based methods. By moving glitch classification into a higher-dimensional feature space, we aim to improve separation between classes exhibiting subtle or overlapping signatures in any particular dimensional plane, thereby contributing to more reliable automated noise classification and detector monitoring in gravitational-wave experiments.

        Speaker: Maria Theodora Folina (Democritus University of Thrace & CERN)
        GRAVCLOUD_CERN.pdf
      • 3:10 PM
        How effectively can Neural Posterior Estimation infer the Neutron Star Equation of State? 1h 10m

        The equation of state (EoS) of neutron star matter encodes the relationship between pressure and density at supranuclear densities, fundamentally governing the star’s structure and observable macroscopic properties, such as mass, radius, and tidal deformability. In this work, we apply Neural Posterior Estimation (NPE) with conditional normalising flows to infer the EoS from synthetic observational data. We consider a model-agnostic EoS family and train our models on mock mass-radius
        and mass-radius–tidal deformability datasets with varying noise levels. We evaluate reconstruction performance in terms of pressure and squared speed of sound across baryonic densities, and
        quantify the impact of including tidal deformability information. Our results demonstrate that tidal
        measurements significantly reduce inference uncertainty, particularly for pressure, and confirm that
        NPE-based models can accurately capture physical constraints.

        Speaker: Valéria Carvalho
        Poster_3.pdf
      • 3:10 PM
        How Many Noise Realizations Do We Really Need? Quantifying Sensitivity Metric Robustness for ML-Based Gravitational-Wave Searches 1h 10m

        We examine the robustness of two common sensitivity metrics for GW detection pipelines, using AresGW model 1, an ML-based detection code. By analyzing the number of detected waveform injections in real detector noise for multiple month-long datasets and the sensitive distance at different false-alarm-ratio (FAR) thresholds, we investigate how these sensitivity metrics fluctuate due to dataset variability. In addition, we evaluate the dependence of these two metrics on contamination by astrophysical GW signals. We also develop a fully calibrated analytical model that explains the observed variance, using Poisson statistics, detector non-stationarity, and the chirp-mass weighting of the sensitive distance. The model identifies threshold variance driven by non-stationarity as the dominant source of variability in the detection count, and provides a first-principles decomposition of the suppression mechanism that makes the sensitive distance a more stable comparison metric. Our findings highlight the challenges introduced by finite-duration datasets and emphasize the need for more rigorous statistical validation. By identifying these challenges, we aim to clarify the practical limitations of both ML-based and traditional detection systems and inform future benchmarking standards for GW searches.

        Speaker: Alexandra Eleni Koloniari (Aristotle University of Thessaloniki)
        How_Many_Noise Realizations.pdf
      • 3:10 PM
        Identifying spacetimes using neural networks 1h 10m

        In general relativity, determining whether two spacetime metric solutions expressed in different gauge describe the same physical scenario poses a significant challenge. This study proposes a novel approach to addressing this problem within the context of numerical relativity by leveraging neural networks. Specifically, we introduce the first implementation of neural networks trained to learn the coordinate mapping between two metric solutions that share identical manifold structure. I will also discuss how this approach could be used to compare various numerical relativity codes, where different choice of gauge and coordinates can affect the final form of the results.

        Speaker: Estuti Shukla (Penn State University)
      • 3:10 PM
        Improved searches for nanohertz gravitational wave background anisotropies with simulation-based inference 1h 10m

        In summer 2023, multiple PTA collaborations reported evidence for a gravitational wave background in the nanohertz regime. Searching for anisotropies is considered a promising way to discriminate between an astrophysical and a cosmological signal origin, but existing methods face key limitations: Bayesian approaches are computationally expensive, while frequentist methods rely on a Gaussian likelihood despite strongly non-Gaussian statistics of the pulsar pair correlations.
        We present a simulation-based, likelihood-free inference framework that replaces the analytic likelihood with a neural network classifier trained on synthetic data. This approach captures the full non-Gaussian structure of the pair correlation estimators and significantly improves performance, more than doubling detection probabilities compared to standard frequentist methods.

        Speaker: Anna-Malin Lemke
        Poster_PTAanisotropies.pdf
      • 3:10 PM
        Isometric embeddings for gravitational wave template banking 1h 10m

        Constructing template banks for matched filtering gravitational wave searches requires placing waveforms densely across a curved parameter space, where distances are defined by a position dependent metric. Evaluating these distances is computationally expensive, and the non-uniform curvature of the space makes uniform template placement suboptimal.

        We propose a neural network approach that learns an approximate isometric embedding of the physical parameter space. In the learned coordinates, Euclidean distances closely approximate true metric distances, reducing template placement to a simple sphere-covering problem in flat space. The network is trained using geometric objectives derived from the induced metric, without relying on density estimation or likelihood-based training.

        Applied to a three dimensional gravitational wave parameter space, the learned embedding achieves ∼98% injection recovery, demonstrating that geometry aware coordinate learning is a promising direction for efficient template bank construction.

        Speaker: Alexandra Wernersson
      • 3:10 PM
        Machine Learning Applications in Numerical Relativity Simulations 1h 10m

        Accurate Gravitational Wave models require data from Numerical Relativity simulations of compact object mergers to inform them of the waveform behaviour during the non-linear merger phase of the binary evolution. Such simulations require large computational resources to reach valuable resolutions, and are time consuming to perform, restricting their ability to fully explore the parameter space of binary mergers. In this talk I will discuss two recent approaches to incorporating Machine Learning techniques into numerical relativity simulations of binary neutron stars; first to accelerate simulation speeds by modelling the nuclear equation of state with neural networks; and secondly to improve the robustness of low resolution simulations by incorporating machine learning models of small scale physics.

        Speaker: William Cook
      • 3:10 PM
        Pinpointing PTA Single Sources: Sequential SBI for Sky Localization 1h 10m

        Accurate localisation of continuous gravitational waves (CGWs) from supermassive black hole binaries (SMBHBs) remains one of the key challenges in Pulsar Timing Array (PTA) data analysis. Traditional searches based on the $\mathcal{F}_e$ statistic provide a robust analytic framework, but the resulting sky maps are strongly affected by the PTA antenna pattern, which redistributes signal power across the sky and generates secondary peaks that complicate the identification of the true source position. This degeneracy motivates the development of alternative approaches capable of disentangling instrumental artefacts from true localisation information.

        In this work, we investigate whether Sequential Simulation-Based Inference (SBI) can improve CGW sky localisation by learning the mapping between $\mathcal{F}_e$ maps and true source positions. Using a suite of simulated PTA datasets, we demonstrate that our SBI pipeline effectively marginalises over antenna pattern search artefacts, providing reliable posterior distributions for the source coordinates. A key advantage of this framework is its extreme computational efficiency: our pipeline can generate $10^5$ $\mathcal{F}_e$ statistic maps in approximately 3 minutes, with the subsequent network training completed in under 2 hours on a single GPU. This enables both the rapid generation of large training sets and near-instantaneous source characterisation. Our results show
        that the angular resolution ($\Delta\Omega$) achieved via SBI is consistent with the theoretical lower bounds predicted by the Fisher Information Matrix.

        Speaker: Ludovica Carbone (University of Milano-Bicocca)
        poster_Ludovica_PTA.pdf
      • 3:10 PM
        Plug-and-Play methods for reconstructing polarizations of gravitational wave signals 1h 10m

        Plug-and-Play methods have been introduced as a generalization of
        variational-based regularization techniques for image-related inverse
        problems. They combine a Gaussian denoising task with a tractable
        likelihood to achieve state-of-the-art image reconstruction. The
        denoising component is flexible and typically implemented using neural
        networks. We present an application of this method to reconstruct the
        polarization of gravitational waves from detector data. A
        proof-of-concept is demonstrated for compact binary coalescences, where
        polarization evolution serves as a signature of relativistic precession.

        Speaker: Thomas SAINRAT
        poster.pdf
      • 3:10 PM
        Simultaneous Misalignment and Mode Mismatch Sensing in Optical Cavities Using Intensity-Only Measurements 1h 10m

        Precise sensing and control of spatial mode content is essential for the performance of precision optical systems, particularly interferometric gravitational-wave detectors, where misalignment and mode mismatch can lead to significant optical losses and degraded quantum noise suppression. Conventional approaches, including heterodyne wavefront sensing and phase camera techniques, are effective but can be limited by hardware complexity and systematic uncertainties arising from restricted reference-beam overlap. This paper presents a novel two-step deep learning pipeline for robust beam diagnostics based solely on beam intensity images. In the first stage, a multi-intensity-image convolutional neural network (CNN) performs accurate mode decomposition, recovering the complex modal content of distorted beams. In the second stage, the predicted mode coefficients are fed into a downstream regression network that simultaneously estimates all eight degrees of freedom (DoFs) associated with misalignment and mode mismatch, including beam tilt, lateral offset, and waist size and position mismatches in both transverse directions. The proposed CNN-based framework achieves a mean absolute error (MAE) of 0.0034 in the mode decomposition stage, which propagates to a total MAE of 0.0062 in the recovered beam imperfection parameters at the final stage. This corresponds to an average residual optical loss of 39 ppm per DoF (310 ppm total). This approach relies only on standard CCD imaging and is robust to random intensity noise, eliminating the need for complex interferometric hardware. The results demonstrate that the proposed deep learning pipeline enables real-time, high-accuracy wavefront sensing and mode-mismatch diagnostics, providing a scalable and hardware-efficient tool for improving the stability and sensitivity of precision optical systems.

        Speaker: Liu Tao (AstroParticule & Cosmologie, Université Paris Cité)
      • 3:10 PM
        The PEARLS project: Precision in Energy-aware AI Research for Low-carbon Solutions 1h 10m

        Artificial intelligence (AI) is increasingly applied in scientific research, but its growing computational demands raise concerns regarding reliability, transparency, and environmental impact.
        The PEARLS project: Precision in Energy-aware AI Research for Low-carbon Solutions, is an ErUM-Data consortium between DESY and the University of Freiburg and aims to address these challenges.
        Within PEARLS, we are developing tools to measure the Carbon footprint of Machine Learning applications, evaluate algorithm efficiency and precision, and improve the processing of data in ML workflows.
        Although PEARLS is being developed within HEP, these approaches naturally extend to other machine learning–based analyses, including gravitational wave physics. This poster introduces the PEARLS work packages and highlights their potential for cross-disciplinary applications.

        Speaker: Josephine Prochnow
      • 3:10 PM
        The Quantum Leap: A Quantum Mechine Learning Approach for Detection of Gravitational Waves in the Context of LISA Space Mission 1h 10m

        Quantum computing and machine learning are two cutting-edge domains, in continuous evolution. Their intersection is quantum machine learning, an area with great potential of enhanced data analysis due to quantum advantage. As gravitational waves astronomy is progressing rapidly, the need for fast and robust data analysis tools increases.
        In our work, we present a quantum neural network data analysis pipeline for detection of gravitational waves signals, in the context of LISA Space Mission. We analyze the mock dataset Sangria v2, consisting of massive black hole binaries signals, provided by the LISA Consortium. We do this analysis in two steps: first, we design and train a variational classifier based quantum neural network for separation of gravitational waves signals from noise. Then, we search for the coalescence times of the signals in the blind dataset. We compare our quantum neural network approach with the classical neural network approach developed in our group, and we highligh the main quantum advantages. Due to current technological limitations, we use for now simulated quantum computer ecosystems.
        In the future, we will addapt our quantum neural network data analysis pipeline to the recent LISA mock dataset, Mojito, and benchmark it against quantum and classical machine learning architectures.

        Speaker: Maria Isfan
      • 3:10 PM
        The Shape of Eccentricity: Rapid Classification of Eccentric Binaries with the Wavelet Scattering Transform 1h 10m

        The gravitational-wave (GW) detections reported by the LIGO-Virgo-KAGRA (LVK) collaboration have so far been consistent with quasi-circular compact binary coalescences (CBCs). Nevertheless, a small fraction of binaries driven to merge through dynamical interactions in dense stellar environments or in field triples may retain measurable orbital eccentricity when entering the sensitive frequency band of LVK detectors. Confident measurement of eccentricity in the LVK band would provide strong evidence for such dynamically driven mergers; however, eccentric gravitational-waveform models are computationally expensive, and performing production-level inference on all detected signals is not an efficient use of resources when eccentric signals are expected to be rare. An intermediate step between detection and analysis, in which the signal is assessed for the potential presence of eccentricity, could provide quick recommendations for which signals should undergo full eccentric inference. We apply the wavelet scattering transform (WST) to a large set of synthetic waveforms in realistic noise and assess its discriminatory power using simple linear and shallow neural-network classifiers. We find that the WST representation enables effective discrimination between eccentric and quasi-circular binaries and provides a compact multi-scale representation of GW signals. Our approach achieves ~64% percent detection accuracy at a false alarm rate of 10%, with an AUC of 0.844 and an average precision of 0.876. We also examine the ability of our classifiers to distinguish eccentricity from spin-induced precession and find robust performance across a range of spin-precession magnitudes.

        Talk based on: https://arxiv.org/abs/2602.24079

        Speaker: Dr Priscilla Canizares (University of Cambridge)
    • 4:20 PM → 5:00 PM
      AI for real-time data processing
      • 4:20 PM
        Machine Learning in LIGO: current and future applications 20m

        The detection of gravitational waves by the Laser Interferometer Gravitational-Wave Observatory (LIGO) has opened a new window onto the universe, but the sensitivity of these detectors is fundamentally limited by a complex and evolving landscape of instrumental and environmental noise. In recent years, machine learning has emerged as a powerful tool for understanding, characterizing, and mitigating these noise sources.
        In this talk, I will present current applications of machine learning in LIGO detector characterization, with a focus on transient noise identification and its impact on search analyses. By leveraging techniques from modern computer vision, including object detection and segmentation models such as YOLO, we are able to automatically identify and localize noise artifacts in time–frequency representations of detector data. These approaches enable scalable classification of noise transients, improved data quality vetting, and more robust separation of astrophysical signals from noise, directly enhancing the sensitivity and reliability of gravitational-wave searches.
        Looking ahead, I will discuss emerging directions where machine learning can play a transformative role beyond data analysis. These include integrating ML into low-latency pipelines, developing physics-informed models for noise prediction and subtraction, and exploring the use of ML-driven optimization in the design of future interferometers. In particular, data-driven approaches may help guide the selection of design parameters for next-generation cavities and instruments, bridging the gap between detector characterization and instrument development.
        Together, these efforts highlight how machine learning is becoming an integral component of both current LIGO operations and the future of gravitational-wave instrumentation.

        Speaker: Mr Siddharth Soni (University of California, Riverside)
        Machine Learning in LIGO Detector Characterization-2.pdf
        Recording
      • 4:40 PM
        Realtime search and inference of binary black holes in LVK data using AI 20m

        The number of GW events have increased from two real-time detections in the LIGO first observing run, to over two hundred in the LIGO-Virgo-KAGRA fourth observing run. In parallel, the last decade has also seen the increased use of machine learning, especially neural networks, in science. For the first time, after a decade of discovery, binary black holes (BBHs) are routinely detected by neural-networks as a part of the LVK data analysis. A total of 23 BBH events were detected in real-time by neural-network based search, Aframe, between late August to mid November 2025. Sky-localization and chirp mass estimates have been distributed in low latency using neural network based parameter estimation algorithm AMPLFI. I will be talking about Aframe and AMPLFI with a view toward this new paradigm of doing low latency GW science using AI.

        Speaker: Deep Chatterjee
        DC-CERN-AI4GW-2026.pdf
        google-slides-link
        Recording
    • 5:00 PM → 6:00 PM
      Tutorials: GW Open Datasets analysis with ML4GW

      https://github.com/TomislavAndric/cern-gw-rl-tutorial

      DLS for GW detectors
  • Thu, May 7
    • 9:30 AM → 10:30 AM
      Synergy with CERN: Contributed Talks
      Convener: Maurizio Pierini (CERN)
    • 10:30 AM → 11:15 AM
      Coffee Break 45m
    • 11:15 AM → 12:15 PM
      Synergy with CERN: Contributed Talks
    • 12:15 PM → 1:30 PM
      Lunch 1h 15m
    • 1:30 PM → 2:50 PM
      AI for data analysis
      Convener: Valerie Domcke (CERN)
      • 1:30 PM
        ★ Linear Oscillatory State-Space Models for Binary Neutron Star Detection ★ 20m

        The detection of long gravitational wave signals in noisy strain data demands models that can efficiently capture long-range temporal structure while remaining computationally tractable. In this talk we introduce Linear Oscillatory State-Space models (LinOSS), a class of sequence models rooted in linear dynamical systems theory, as an alternative to conventional deep learning architectures for gravitational wave data analysis. LinOSS processes time series through a learned latent state that evolves via structured recurrence, incorporating an oscillatory inductive bias that reflects the structure of the underlying physical systems. This enables the model to capture long-duration dependencies while scaling compute time logarithmically with sequence length. We then explore the application of this framework to the classification problem of distinguishing binary neutron star (BNS) merger signals from detector noise.

        Speaker: Benedict Armstrong
        Recording
        slides.pdf
      • 1:50 PM
        ★ From Inspiral to Inference: BNS Parameter Estimation with State Space Models ★ 20m

        Fast and accurate parameter estimation of binary neutron star (BNS) mergers, gravitational wave events with electromagnetic counterparts, remains a central challenge in multimessenger astronomy. Building on the State Space Model (SSM) framework presented in the companion talk, we directly regress BNS merger source parameters from raw gravitational wave time series, without sampling-based inference. As a first demonstration, we focus on the chirp mass, the dominant parameter governing the inspiral waveform. We show that regression succeeds not only from windows centered on the merger, but also from pre-merger inspiral segments, with implications for early-warning detection pipelines. By placing a Gaussian prior on the regression target, we additionally regress uncertainty estimates on the inferred chirp mass. We further demonstrate the physical meaningfulness of these uncertainties by showing they discriminate between signal and background events with performance comparable to existing pipelines. We discuss prospects for extending this framework to additional source parameters, including sky localization, and explore potential improvements through preprocessing strategies such as denoising.

        Speaker: Kyungseop Yoon (Massachusetts Institute of Technology)
        BNS_PE_SSM_(KYoon).pdf
        Recording
        skymap_50_events.gif
      • 2:10 PM
        ★ SlotFlow: Amortized Trans-Dimensional Inference towards LISA ★ 20m

        Gravitational-wave observations from future space-borne detectors will present a fundamentally new inference challenge: not only we estimate the parameters of each source, but we must simultaneously determine how many sources are present. This is the trans-dimensional Bayesian inference problem, and classical approaches such as Reversible Jump MCMC can take hours to days per analysis.

        We present SlotFlow, a deep-learning architecture for amortized trans-dimensional inference. Given a time-series observation, SlotFlow jointly infers the number of signal components $K$ and their individual parameters in a single forward pass, achieving millisecond-scale posterior estimation with well-calibrated uncertainties. The architecture combines (i) a dual-stream frequency–time encoder that extracts complementary spectral and temporal representations, (ii) a classifier that predicts the cardinality posterior from spectral features, and (iii) dynamic slot allocation that instantiates exactly $K$ slot contexts, each passed to a single shared conditional normalizing flow that produces per-source posteriors. Training uses permutation-invariant Hungarian matching to handle the inherent label-switching symmetry of multi-source posteriors.

        We validate SlotFlow on sinusoidal mixtures with up to 10 overlapping components, a canonical benchmark motivated directly by the quasi-monochromatic nature of LISA galactic binaries. SlotFlow achieves high cardinality accuracy and posteriors in close agreement with RJMCMC across amplitude, phase, and frequency parameters, while reducing inference time from hours to milliseconds. Posterior calibration remains reliable across the full range of source counts and signal-to-noise ratios explored.

        Speaker: Giovanni Giarda
        CERN_SlotFlow.pdf
        Recording
        SlotFlow_CERN.pptx
      • 2:30 PM
        ★ GWAK2: Gravitational Wave Anomalous Knowledge using SSL ★ 20m

        Since the first gravitational-wave detection by ground-based interferometers, after more than a decade of observations has yielded over one hundred compact binary coalescence (CBC) events, whose waveforms can be well-modeled by general relativity. These well-modeled signals enable detection pipelines based on matched filtering, which search for waveform consistency against the CBC template bank. However, within the sensitivity band of the LIGO–Virgo–KAGRA Collaboration detectors, a broader class of astrophysical sources such as core-collapse supernovae and other transient phenomena are poorly modeled or inherently unpredictable. Consequently, these sources cannot be efficiently captured by template-based searches, motivating the need for waveform-agnostic detection strategies. To address this challenge, we develop Gravitational Wave Anomalous Knowledge (GWAK), a machine learning–based search framework designed for generic transient detection. GWAK employs a semi-supervised embedding model using Self Supervised Learning to learn a low-dimensional representation of detector data, followed by a metric model trained on noise data to define a discriminative search space. This approach enables waveform-agnostic searches for gravitational-wave transients and improves the identification and characterization of unexpected signals.

        Speakers: Andy Chen (Institute of Physics, National Yang-Ming Chiao Tung University, Hsinchu, Taiwan), Eric Anton Moreno (Massachusetts Institute of Technology (US))
        GWAK2: Gravitational Wave Anomalous Knowledge using SSL
        GWAK2: Gravitational Wave Anomalous Knowledge using SSL
        Recording
    • 2:50 PM → 3:20 PM
      Coffee break 30m
    • 3:20 PM → 5:00 PM
      AI for detector operations
      Convener: Jan Harms (INFN - National Institute for Nuclear Physics)
      • 3:20 PM
        From Post Hoc Subtraction to Source Suppression: ML Noise Mitigation in Advanced LIGO 20m

        We present our work on machine learning for noise mitigation in Advanced LIGO that moves from software denoising of the strain channel to suppression of disturbances at their source within the detector control system.
        Using Coherence DeepClean, we perform coherence-based witness-channel selection followed by machine-learning regression to subtract linear and nonlinear noise couplings from the gravitational-wave readout after they enter the strain data. This software denoising approach has yielded measurable sensitivity gains, including a 4.3% improvement in astrophysical sensitive volume.
        We extend the same data-driven philosophy upstream to the seismic isolation system, where neural-network models predict residual platform motion induced by persistent microseismic activity in the 0.1–0.3 Hz band. In contrast to post hoc subtraction, this method targets the disturbance before it propagates through the instrument, enabling direct suppression of motion at the source. Our results suggest that, if integrated into the control system, the method could offer up to an order-of-magnitude reduction in residual motion compared to conventional linear filtering.
        Looking ahead, we are exploring reinforcement learning for increasingly autonomous control architectures. Together, our results outline a path toward autonomous machine-learning systems for improving detector stability, low-frequency sensitivity, and astrophysical reach.

        Speaker: Christina Reissel (Massachusetts Inst. of Technology (US))
        Denoising_AI4GW_CERN_CReissel.pdf
        Recording
      • 3:40 PM
        Non-linear noise regression in the Virgo detector 20m

        Since the first detection of gravitational waves from a binary black hole merger, hundreds of such events have been observed. However, many compact binary coalescence signals remain buried below the detector noise threshold and could be recovered through improved noise mitigation. Detector noise arises from multiple sources, including fundamental, technical, and environmental contributions (seismic activity, atmospheric fluctuations, thermal variations, and Newtonian noise).
        These sources can be stationary, non-stationary, or non-linear; the latter generate complex and unpredictable patterns that make signal recovery particularly challenging. This work presents the first demonstration of non-linear noise regression using deep learning techniques within Virgo. Using DeepClean, a convolutional neural network, both stationary and non-stationary noise contributions are modeled and subtracted from strain data, revealing previously unmodeled non-linear noise couplings and improving detector sensitivity.
        The method leads to measurable performance gains, including an increase of up to 1.3 Mpc (~2.5%) in the binary neutron star inspiral range and a ~1.7% average improvement in recovered signal-to-noise ratio for injected binary black hole signals. These results demonstrate the potential of machine learning-based noise mitigation for enhancing gravitational-wave detection and enabling low-latency transient searches.

        Speaker: R. Weizmann Kiendrebeogo (IRFU, CEA, Universit´e Paris-Saclay, F-91191 Gif-sur-Yvette, France)
        Weizmann_CERN_AI_for_GW_Workshop.pdf
      • 4:00 PM
        Tools for real-time inferences in GW detectors 20m

        Machine learning will play a key role in the next generation of interferometers data acquisition architectures, particularly through hardware-based solutions. However, methods deployed will have to meet very specific simplicity and robustness requirements. We will present those constraints and the tools we are currently developing to fulfill them at different hardware stages:

        • TolmNet: a C program included in the Virgo online software framework to perform neural network inference in real-time computing environment.
        • pyML2FPGA: a python framework based on an HDL library converting a network into a ressources/latency optimized VDHL code which can be ported to a low-end FPGA.

        In both cases, we present the results obtained so far using as an input a simple gravitational wave detection network.

        Speaker: Sebastien Viret (Centre National de la Recherche Scientifique (FR))
        AIGW_060526_v3.pdf
        Recording
      • 4:20 PM
        ★ Computational Discovery of Interferometric Gravitational Wave Detectors ★ 20m

        Current and next-generation gravitational wave detectors are designed by human experts who must balance coupled physical effects across many domains. The vast space of all possible experiment designs suggests that many high-sensitivity, unconventional detectors may lie beyond the reach of human intuition alone. AI-based methods are increasingly capable of discovering powerful measurement schemes from first principles, offering a complementary design paradigm with biases distinct from those of human experts. We therefore frame the discovery of novel gravitational wave measurement techniques as a search for optima over a vast space of hardware configurations subject to practical constraints. We discuss how to engineer an expressive search space with the potential to discover novel detector topologies and present Differometor, a differentiable interferometer simulator built for high-performance optimization. We then formulate gravitational wave detector design as a challenging algorithmic benchmark and argue that new interpretability and analysis tools will be essential for understanding and exploiting unconventional AI-discovered detector blueprints.

        Speaker: Jonathan Klimesch (University of Tübingen)
        26.05.07_CERN.pdf
        Recording
      • 4:40 PM
        Deep Loop Shaping for Angular Sensing and Control in Virgo 20m

        Gravitational-wave interferometers rely on hundreds of feedback control loops to stabilize mirror alignment and maintain detector sensitivity. These control systems can inject noise into the observation band, limiting low-frequency performance. Recently, the Deep Loop Shaping (DLS) approach demonstrated that reinforcement learning (RL) can substantially reduce injected control noise in the 10–30 Hz band in the LIGO Livingston Observatory, achieving up to two orders of magnitude suppression on a key alignment loop.

        Motivated by these results, we investigate the application of RL-based loop shaping to the Virgo detector. We developed a Virgo-specific Lightsaber time-domain simulation of the DIFFp_TX angular control loop and implemented a distributed RL training pipeline. The environment includes realistic plant dynamics, sensing noise, and stabilization through a baseline linear controller, while the RL agent learns control policies using frequency-domain reward functions designed to suppress control noise while preserving loop stability and taming radiation-pressure effects.

        We present the architecture of the training framework, the Virgo Lightsaber model, and results comparing key observables between RL-assisted control and the baseline linear controller. This work represents a crucial step towards the first deployment of reinforcement-learning-based control policies in the Virgo GW detector.

        Speaker: Tomislav Andric (Gran Sasso Science Institute & INFN-LNGS)
        DLS_AI4GW_CERN_TAndric.pdf
        Recording
    • 5:00 PM → 6:00 PM
      Tutorials: Training of a reinforcement-learning controller using a Virgo or LIGO digital twin

      https://github.com/TomislavAndric/cern-gw-rl-tutorial

      DLS for GW detectors
  • Fri, May 8
    • 9:00 AM → 10:40 AM
      AI for real-time data processing: Contributed Talks
      • 9:00 AM
        ★ Scalable detection of long-duration gravitational wave signals from neutron star binaries in next-generation ground-based detectors ★ 20m

        Gravitational waves, ripples in the fabric of spacetime produced by accelerating cosmic masses, are routinely detected by the ground-based LIGO-Virgo-KAGRA (LVK) network. The historic first observation of a Binary Neutron Star (BNS) coalescence, GW170817, was tracked in the detectors’ sensitive band for about tens of seconds. This observable duration is dictated by the interferometers’ minimum frequency limit. While current-generation facilities are restricted by seismic noise to frequencies above ≈ 20 Hz, the advent of Third-Generation (3G) observatories, such as the Einstein Telescope (ET) or the Cosmic Explorer (CE), will push this boundary down to the few-Hertz regime, enabling the observation of BNS coalescences for several hours before merger. Over these extended observation windows, the standard assumption of a static detector breaks down. The Earth’s rotation and orbital motion induce amplitude and frequency modulations such as Doppler shifts and antenna pattern variations. Tracking these over long durations requires unaffordable computing costs, limiting the scalability and low-latency applicability of classical algorithms, such as Matched Filtering.

        To reduce this computational burden, we propose marginalizing over the extrinsic parameters using machine learning, restricting the pipeline to a mass-only search. To achieve this, we design a 2D Convolutional Neural Network (CNN) that processes multi-detector, complex Short Fourier Transform (SFT) tensors. By optimizing the Binary Cross-Entropy (BCE) loss, the network approximates the optimal Bayesian detection statistic, allowing the CNN to learn the non-linear time-frequency morphology of the signal and implicitly performing the required marginalization during training. We conduct a large-scale campaign of software-simulated signals to compare the CNN’s performance with a GPU-optimized SFT-based Matched Filtering baseline across three signal duration regimes: 𝑇sig ∼ 25 minutes (𝑓min = 5 Hz), ∼ 1.88 hours (3 Hz), and ∼ 6 hours (2 Hz).

        We found that the CNN achieves a detection sensitivity comparable to the Matched Filtering baseline, while reducing the overall computational cost by over four orders of magnitude and slashing inference latency from minutes to milliseconds. This gain in efficiency enables a low-latency triggering of long-duration gravitational wave signals, suggesting Machine Learning as a crucial tool for earlywarning multi-messenger astronomy in the upcoming 3G era.

        Speaker: Martin Gerini (UCLouvain)
        MARTIN_GERINI_AI4GW.pdf
        MARTIN_GERINI_AI4GW.pptx
        Recording
        Scalable detection.pptx
      • 9:20 AM
        ★ BOAW: a neural network approach to combining multiple pipelines in gravitational-wave searches ★ 20m

        Gravitational-wave transient searches in LIGO-Virgo-KAGRA (LVK) routinely run multiple low-latency (and offline) pipelines in parallel. Their redundancy and complementarity improves robustness, but it also creates a practical challenge: how to combine pipeline outputs into a single, reliable detection statistic without resorting to mere union of their findings. We present BOAW (Best-Of-All-Worlds), a machine-learning meta-pipeline that fuses evidence from four Compact Binary Coalescence search pipelines (GstLAL, MBTA, PyCBC, and cWB-BBH) used in the LVK for real-time astronomical alert generation, but also in offline catalogs of gravitational-wave detections. BOAW ingests a compact, interpretable feature set from each pipeline and treats non-triggers (from any of the aforementioned pipelines) as informative rather than missing data. We use an ensemble of five diverse neural-network architectures, selected through neural architecture search and trained on LVK mock-data-challenge injections and real instrument noise in order to produce a unified detection statistic. On held-out data, BOAW achieves an AUC of 0.99, delivering up to an order-of-magnitude reduction in false-positive rate at 90% detection efficiency compared to individual pipelines and their logical combinations. The overall algorithmic architecture is flexible, computationally light-weight for training and can be deployed for real-time public alert generation as well as for offline analysis (event catalog generation) without adding any measurable latency. We propose its adoption by the LVK for its ongoing observations.

        Speakers: Nikolas Moustakidis (Aristotle University of Thessaloniki), Mr Theofilos Moustakidis (University of Thessaly)
        NM-TM-CERN-AI4GW-2026_v3.pdf
        NM-TM-CERN-AI4GW-2026_v3.pdf
        NM-TM-CERN-AI4GW-2026_v3.pptx
        Recording
      • 9:40 AM
        ★ Normalizing flows for complete parameter estimation on time-frequency representations of gravitational-wave data ★ 20m

        The speed-up of parameter estimation is an active field of research in gravitational-wave data analysis. In this work we present GP15, a deep-learning method that merges residual networks and normalizing flows into a general-purpose, image-based estimator of binary black hole (BBH) parameters. Building on our early work, we map BBH spectrograms from the Advanced LIGO and Advanced Virgo detectors to color channels in an RGB image amenable to be processed with residual networks. GP15 is trained on simulated data for BBH mergers obtained with the \texttt{IMRPhenomXPHM} waveform approximant and tested for all three-detector events from the GWTC-3 and GWTC-2.1 catalogs reported by the LIGO-Virgo-KAGRA (LVK) collaboration. Overall, our model yields good agreement with the LVK results over most parameters. Our simple model can produce large amounts of posterior samples in the order of a second, complementing existing approaches with normalizing flows based on time-only or frequency-only representation of gravitational-wave data.

        Speaker: Daniel Lanchares (Universidad de Oviedo - ICTEA)
        AIGW_2026_Slides_Daniel_Lanchares.pdf
        Recording
      • 10:00 AM
        GWEEP: A Deep Learning Toolkit for Low‑Latency Gravitational‑Wave Analysis 20m

        With the LISA mission formally adopted by ESA in January 2024 and now in its implementation phase, preparations across the science ground segment are accelerating toward launch in the mid‑2030s. A central priority is ensuring that data‑analysis tools are ready to extract science quickly and reliably once telemetry becomes available.
        Within this context, here we present GWEEP(Gravitational Wave DEEp-learning Pipeline), a deep‑learning toolkit designed for rapid detection and parameter estimation of gravitational‑wave signals.
        GWEEP combines efficient neural architectures with domain‑specific pre‑processing to operate on streaming batches, enabling low‑latency triage of candidate transients and early characterisation of their source parameters.
        We illustrate the pipeline design and summarise performance on recent LISA‑like datasets. For validation purposes we’ve used the LISA data challenges, a set of realistic LISA mock data prepared and launched periodically by LISA LDC group that offered the perfect environment to develop and test the data processing tools within the Consortium.
        We conclude by outlining the roadmap for deployment within the consortium’s data‑processing ecosystem.

        Speaker: Ana Caramete (Institute of Space Science - INFLPR Subsidiary)
        GWEEP_CERN_0526_ACv2.pdf
        GWEEP_CERN_0526_ACv2.pptx
        Recording
      • 10:20 AM
        The Present and the Future of Machine Learning for Multi-Messenger Astrophysics 20m

        With the detection of compact binary coalescences and their
        electromagnetic counterparts by gravitational-wave detectors, a new
        era of multi-messenger astronomy has begun. In this talk, I will
        describe how machine learning is enabling the gravitational-wave community to make very low-latency detection and parameter estimation possible within the alert system. I will then discuss how current ground based optical surveys and dedicated follow-up systems are integrating machine learning into their standard work flows, with examples from both current and near future surveys. We will close with near-term prospects for the field.

        Speaker: Michael Coughlin (University of Minnesota)
        CERN_MMA.pdf
        CERN_MMA.pdf
        CERN_MMA.pptx
        Recording
    • 10:40 AM → 11:30 AM
      Coffee Break 50m
    • 11:30 AM → 12:30 PM
      AI for detector operations
      • 11:30 AM
        GlitchFlow: generative neural networks for de-noising gravitaional waves data in present and future detectors. 20m

        Third-generation (3G) gravitational waves (GW) detectors such as Einstein Telescope and Cosmic Explorer are expected to increase annual GW detections by a factor of 1000, enabling the detection of every stellar-mass black hole merger in the Universe. The growing volume and complexity of data generated by these observatories necessitate the application of data science techniques, particularly deep learning, to facilitate discoveries. One of the main challenges that 3G detectors will face is the reduction of noise. Specifically, the sensitivity will be most effected by the presence of transient noise in the form of glitches appearing in the detector data at much higeher rates than for previous interferometers. For this reason noise reduction and subtraction is one of the most important activities in GW research.

        We present GlitchFlow, a novel Machine Learning pipeline for vetoing and de-noising on GW interferometer data. GlitchFlow uses a Generative Neural Network with a complex U-Net architecture to map carefully selected auxiliary channels (sensitive to the noise but not to GW signals) into the main channel of the interferometer. The generated output can then be subtracted from the interferometer main channel data. We achieve a de-noising accuracy above 90% for glitches of SNR 6 or higher on data from the Advanced Virgo detector. The input and ouput data consist of 2D spectrograms computed using the newly developed Q-Transform Amplitude Modulation (QTAM) package, which encodes all the physical information of the signals into amplitude and phase maps. This innovation makes it possible invert the spectrograms and obtain the cleaned data as a 1D time series, opening the path for hybrid 1D-2D analysis workflows and providing an output which can be directly fed into the standard Parameter Estimation pipelines.

        Speaker: Francesco Sarandrea
        GlitchFlow_AI4GW_CERN_2026.pdf
        Recording
      • 11:50 AM
        Operations of GW alert validation in O4 20m

        In my talk, I would like to discuss the role of large language models (LLMs) in supporting GW candidate validation during O5 observing run, in complement to shifts rota that were in place to provide reliability to the follow-up and mitigate the risk to consume telescope ressources for no reason. While human-in-the-loop validation has remained critical over the last campaigns, we are raising new challenges with the increasing rates of alerts and their diversity. It was also hard to maintain during O4 a certain number of participation in the O4 operations. In my talk, I will explore pro/cons LLM-based framework trained on past operational knowledge, metadata, and expert interactions to assist operations in the alert dissemination and reduce the shift workload. I will also enumerate the key ingredients to make it come true which also include the connection of data science (that can be related to same problematics than CERN for operation in data analysis).

        Speaker: Sarah ANTIER (IJCLAB)
        Recording
        RRT_O5_Antier.pdf
      • 12:10 PM
        Deployable Reinforcement Learning for Interferometric Control 20m

        Gravitational-wave detectors present a difficult control problem. They are strongly coupled, nonstationary, and safety-critical, and there is very little room to learn directly on real hardware. In this talk, I will describe a practical approach to reinforcement learning for interferometric sensing and control, drawing on lessons from both kilometer-scale interferometers and tabletop optical testbeds. I will focus on building useful observations from the underlying physics, using scalable simulations and reduced-order models to make training practical, reuse of existing hardware data, and adding supervisory checks for safer online operation. The broader aim is not to replace classical control, but to combine it with search and learning in a way that is more adaptive and more data-efficient for precision optical systems.

        Speaker: Nikhil Mukund (MIT)
        Deployable_RL_for_Interferometric_Control_CERN_May2026_Nikhil.pdf
        Recording
    • 12:30 PM → 12:50 PM
      Student Prize Ceremony
      BestAwards_AI4GW.pdf