GDR Ondes Gravitationnelles : Rencontre du groupe de travail "méthodes d'analyse des données"

mercredi 7 oct. 2026, 08:55 → 17:45 Europe/Paris

Dirac/RdC-Amphithéâtre (IP2I)

Cet événement est organisée le 7 octobre, en satellite de la huitième assemblée générale du Groupement de recherche ondes Gravitationnelles (5-6 octobre, agenda).

Le GdR Ondes Gravitationnelles (http://gdrgw.in2p3.fr/) a été crée en 2017 avec le but de rassembler la communauté scientifique intéressée par l’exploration de l’Univers avec les ondes gravitationnelles, et de lui fournir des occasions de rencontres et de discussions communes.  Le groupe de travail Méthodes d'analyse des données est dédié aux techniques de traitement et analyse des données et encourage les synergies, dans ce domaine important, entre différentes expériences et communautés.

L'inscription à cette journée est obligatoire avant le 5 septembre 2026.   

La date limite pour proposer une contribution est le 5 septembre 2026.    

La matinée sera consacrée à des présentations d’experts sur l’utilisation des outils d’intelligence artificielle appliqués à la recherche d’ondes gravitationnelles.

L’après-midi sera consacrée à des présentations plus courtes sur des sujets variés. Toutes les thématiques du groupe de travail sont les bienvenues.

Connexion zoom 

08:55 → 09:00 Introduction

Orateurs: Alba ROMERO RODRIGUEZ (Laboratoire D'Annecy de Physique de Particules (LAPP)), Florian Aubin (IPHC), Yves Lemiere (LPCCaen, ENSICAEN, Unuiversité de Caen, CNRS/IN2P3, Caen, France)

09:00 → 10:30 Session 1: Artificial intelligence tools used in gravitational wave data analysis

09:00 Machine Learning for Gravitational Wave Data Analysis

After a brief overview of Machine Learning fundamentals, this presentation offers a panorama of the main ML methods applied to gravitational-wave data processing: Simulation-Based Inference (SBI), transformer models, and diffusion models.

This panorama will be illustrated by the MANGO project, which is developing an ML framework to address the challenges posed by future detectors (Einstein Telescope and LISA), whose expected detection rates will exceed the capabilities of traditional MCMC methods.

The aim is to give the community a shared foundation of understanding, as an introduction to the more specific presentations of the day.

Orateur: Antsa Rasamoela (L2I Toulouse, CNRS/IN2P3, Université de Toulouse)

MANGO_GDR_OG

09:45 Detection of GW stochastic background with Machine Learning

Orateur: Hugo Einsle (Universiteit Antwerpen)

Hugo_EInsle_GdR.pdf

10:30 → 11:00 Coffee break

11:00 → 12:30 Session 2: Artificial intelligence tools used in gravitational wave data analysis

11:00 Search for supernovae with ML

Orateur: Alberto Iess (LAPP CNRS)

ML_CCSN_GDR_OG_Alberto_Iess.pdf

11:45 Combining matched filtering with a convolutional neural network for CBC signal detection

When searching for compact binary mergers using matched filtering, we usually generate candidate events by finding the local maxima of the signal-to-noise ratio as a function of the search parameters, applying a threshold to them, and then forming a ranking statistic that combines the signal-to-noise ratio with signal-based discriminators, to account for nonstationary and non-Gaussian noise. I will present a proof-of-concept study that uses matched filtering in a different way, namely feeding the ensemble of its results into a convolutional neural network. This leads directly to a ranking statistic that accounts for the collective behavior of the signal-to-noise ratio over a given window of the search space. Using simulated noise and signals, I will show the behavior of this ranking statistic and compare it to conventional quantities like the maximum signal-to-noise ratio and the chi^2-weighted maximum signal-to-noise ratio.

Orateur: Tito DAL CANTON (UPSaclay, CNRS/IN2P3, IJCLab)

MF and CNN for CBCs.pdf

12:30 → 14:00 Lunch

14:00 → 15:45 Session 3: Contributed talks (ML)

14:00 Accurate and efficient simulation-based inference for massive black-hole binaries with LISA

We develop an accurate simulation-based inference framework for high-mass (≳ 10⁷M⊙) 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 at fixed reference time. After sampling, we importance-sample to the true posterior based on the underlying likelihood and prior. 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 proposed 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.

Orateur: Alice Spadaro (L2I Toulouse, CNRS/IN2P3, Université de Toulouse)

ASpadaro_GDR_Lyon.pdf

14:20 A simulation-based inference pipeline for fast parameter estimation of massive black hole binaries with LISA data

This work focuses on data analysis for LISA, with a particular focus on fast parameter inference of massive black hole binaries (MBHBs). Fast parameter estimation will be essential for LISA’s Low-Latency Alert Pipeline (LLAP), which will enable electromagnetic and multi-messenger follow-up of detected sources. This requires working with pre-merger signals, allowing other observatories to see the detected events thanks to the alerts given by the pipeline. These constraints and the high computational cost of standard Bayesian methods motivate the use of amortized approaches such as simulation-based inference (SBI).

We present a modular SBI pipeline for MBHB parameter estimation using LISA time-domain data. It includes dataset generation with a fast response that we developed, training of a neural posterior estimator, refinement of the sampling results with importance sampling, and posterior validation. The analysis is performed directly on time-domain signals to avoid pre-processing of the received data. The modularity of the pipeline allows us to independently test and improve each of its components.

We explore several approaches to improve the accuracy and efficiency of the inference, including learned summary statistics to compress the time-domain signals, restricting the priors based on a detection signal-to-noise ratio threshold, and testing other models such as flow matching. We also work on global standardization of the parameters instead of batch standardization, in order to better handle the large range of parameter values.

Orateur: Louis Le Saulnier (CEA/IRFU/DEDIP)

GdR_2026_LLS.pdf

14:40 Low-latency detection and parameter estimation of massive black hole binaries for LISA

LISA will observe massive black hole binaries (MBHBs) hours to weeks before they merge. If these systems are in gas-rich environments, they might produce an electromagnetic (EM) counterpart: this motivates the development of pipelines able to localize and characterize the source fast enough to trigger a follow-up search. The main data analysis pipeline, the Global Fit, will model all LISA sources jointly, which makes it too slow for near-real-time alerts. I will present two approaches for quick detection and parameter estimation of MBHBs.
The first is a simulation-based inference approach that employs conditional flow matching with optimal transport: we train a network to learn the trajectory from a base distribution to the parameters' posteriors, conditioning on data. The computational cost is shifted to the training phase, allowing the trained model to produce posteriors within low-latency requirements. We test the approach on both time and time-frequency representations of the data, and project our data into a space where other sources are ignored, by training autoencoders to remove the galactic binary foreground.
The second, in collaboration with Aryamann Rao (UBC), is a detection pipeline that could provide starting points for deeper analyses: through excess power in the Q-transform, a time-frequency representation of the data, we identify candidate signals. We give a first estimate of chirp mass and coalescence time and, based on this trigger, we select which network to run between several trained CNNs, each specialized on a restricted mass range. We provide a quick point estimate of the full set of parameters. We have used this approach to detect sources in the Mojito Light data challenge, and we are now testing the CNNs performance on those detections.

Orateur: Malvina Bellotti (Observatoire de la Côte d'Azur, Nice)

GraviGammaNu - GDR Lyon.pdf

15:00 Diffusing the assumptions: realistic LISA stochastic component separation with physics-aware simulation-based inference

The numerous sources that LISA will observe pose a unique data analysis challenge, with the most promising current avenue being a Global Fit, a simultaneous inference of all sources. Within the LISA Global fit, stochastic sources, from instrumental noise, to the galactic foreground and astrophysical or cosmological backgrounds, are particularly relevant due to their effect on the quality of the inference of all other deterministic sources. Issues such as non-gaussianity, due to poor subtraction of deterministic signals, or data gaps, scheduled and unscheduled, provide a unique opportunity for machine learning to augment traditional Bayesian methods. We present a stochastic signal inference framework for LISA data analysis, based on simulation-based inference, with a state-of-the-art all-in-one framework, the Simformer. We further augment the inference network with a novel physics-aware embedding scheme. This framework can recover the individual stochastic components, including spline-based and per-link noise degrees of freedom, in the presence of realistic glitch rates and antenna pattern-repointing-driven gaps or a supermassive black hole binary. Furthermore, we provide a comprehensive testing suite of our network, both at a population and per-event level.

Orateur: Iuliu Cuceu (Observatoire de la Côte d'Azur)

Cuceu_GDR_2026-10-07_v2.pptx

15:20 Plug-and-Play methods for reconstructing polarizations of gravitational wave signals

Reconstructing gravitational wave polarizations from detector data is a challenging problem due to the low SNR and the ill-posed nature of the the problem. We leverage the so-called "Plug-and-Play" methods, commonly used in inverse problem literature to achieve state-of-the-art image reconstruction. They explicitely separate the differentiable likelihood and the regularization, which is likened to a Gaussian denoising task. This denoising tasks can be achieved in a variety of ways, typically using a neural network. We demonstrate the usage of this method on compact binary coalescences, allowing to obtain a direct measurement of relativistic precession.

Orateur: Thomas Sainrat (CentraleSupélec)

Presentation GdR 2026 SAINRAT.pdf

15:45 → 16:15 Coffee break

16:15 → 17:00 Session 4: Contributed talks

16:15 Fast localization of massive black hole binaries using the coronagraphic method for the LISA mission

The LISA constellation offers a unique opportunity to detect Massive Black Hole Binaries (MBHBs), with a sensitivity band that is particularly well suited to these sources. Their detection will allow us to trace the origins, growth and merger histories of massive black holes, to explore the fundamental nature of gravity and black holes and to probe the rate of expansion of the Universe using these sources as standard sirens [Colpi, 2024].

To fully exploit MBHB observations, rapid and reliable sky localization is crucial [Mangiagli, 2022], as it enables the triggering of alerts for electromagnetic observatories through the Low-Latency Alert Pipeline (LLAP), complementing the observation of these complex systems with a multi-messenger approach. In this context, we present a fast and agnostic localization method.

This approach relies on the construction of an observable, the coronagraphic variable κ derived from linear combinations of Sagnac-type channels [Costa, 2025]. The principle is as follows : when the direction and frequency parameters injected into the observable exactly match the true source parameters, the gravitational-wave signal is canceled within the κ variable.
Localization is thus performed by scanning the celestial sphere to identify the configuration that minimizes the spectral power of the signal.

Preliminary results from the implementation of this approach within the NullStream pipeline have already proven promising for MBHBs [Costa, 2025]. We are now working to improve the accuracy of this localization. Moreover, since the method is fast owing to its agnostic nature, it is naturally well suited to integration within the Low-Latency Alert Pipeline (LLAP).

Finally, as this remains to be investigated, one may ask how the sky localization obtained through the Coronagraphic method could later be used to constrain full Bayesian parameter estimation.

Orateur: Hugo Alexandre (LPC Caen)

Hugo_Alexandre-GDR2026.pdf

16:35 Earth's rotation impact on long-duration signals for 3G detectors

We studied the impact of the Earth's rotation on BNS signals that would be detected by Einstein Telescope (~ 2 hours).
Especially, we studied how the Match and the SNR would vary if we take (or not for simplicity) into account the modulations induced by the Earth's rotation on the templates

We show the results for the 2L and Xylophone configurations, both for a starting detection frequency of 2Hz and 5Hz, and included in some 3G detectors network configurations.

Orateur: Mathieu Schoor (LAPP)

GdR Lyon 102026-2.pdf

Lien diapo