IA@IRAP 2025 Day

mercredi 15 oct. 2025, 09:00 → 18:00 Europe/Paris

IRAP (Toulouse)

(English below)

Introduction

Au cours des dernières décennies, les techniques d'apprentissage automatique, le Machine Learning (ML), le Deep Learning (DL) et l’Intelligence Artificielle (IA) en général ont suscité un grand intérêt dans notre communauté. En effet, les instruments de plus en plus performants, déjà opérationnels sur les télescopes actuels ou ceux en cours de construction (notamment dans le cadre de différents projets portés par l’IRAP), fournissent et fourniront des quantités incroyables de données, en particulier dans le cas des multiples relevés d’objets astrophysiques réalisés. Cette réalité génère un besoin urgent de mettre en œuvre des méthodes et des algorithmes capables d'exploiter ces données, de les classer, de les analyser et d'en extraire la physique sous-jacente.

Ce phénomène n’a pas échappé à notre laboratoire. Plusieurs membres de l’IRAP se sont intéressés de près ou de loin à ces méthodes, ce qui a motivé la création du groupe ML/DL de l’IRAP en 2022.

L’utilisation de l’IA au sein de l’IRAP concerne un large domaine de thématiques scientifiques représentées par les équipes du laboratoire. Elle implique différentes approches d’apprentissage automatique, qu'il s'agisse d'approches supervisées ou non supervisées, de classification ou de régression.

Cette journée « IA@IRAP » a pour objectif de réunir les membres de l’IRAP intéressés par ces méthodes, voulant approfondir leurs connaissances dans le domaine ou travaillant avec ces méthodes sur des thématiques et des données astrophysiques, afin de proposer un espace d’interaction et de retour d’expérience et de créer une dynamique autour des applications de l’IA dans le domaine de l’astrophysique, des mesures et de l’instrumentation.

Nous aurons deux experts invités, Jonathan Sprauel (expert en IA de l’IRT Saint-Exupéry, directeur opérationnel ANITI) qui fera une présentation générale d’introduction aux méthodes, approches et limites de l’IA et Karin Dassas (Membre du bureau d’Ecoinfo, CNRS, CESBIO) qui nous parlera de l’impact écologique des technologies du numérique en général.

Une table ronde est prévue lors de cette journée à laquelle vous êtes conviés de participer.

Comité d'organisation :

• Hui Yang, chercheur postdoctoral, GAHEC, IRAP

• Rungployphan Kieokaew, Ingénieure en IA et physique spatiale, IRAP et Inria.

• Jihane Moultaka, astronome-adjoint, GAHEC, IRAP

------------------------------------------

Introduction

Over the past few decades, Machine Learning (ML), Deep Learning (DL) and Artificial Intelligence (AI) techniques in general have instigated great interest in our community. Indeed, the increasingly high-performance instruments already operational on current telescopes or those under construction (notably as part of various projects supported by IRAP) are providing, and will continue to provide, massive amount of data, particularly in the case of the multiple surveys of astrophysical objects being carried out. This reality generates an urgent need to implement methods and algorithms capable of exploiting this data, classifying it, analyzing it and extracting the underlying physics.

Our laboratory has taken note of this phenomenon. Several IRAP members have shown a keen interest in these methods, which led to the creation of the IRAP ML/DL group in 2022.

The use of AI at IRAP concerns a wide range of scientific themes represented by the laboratory's teams. It involves various machine learning approaches, whether supervised or unsupervised, classification or regression.

The aim of this “IA@IRAP” day is to bring together IRAP members who are interested in these methods, wishing to deepen their knowledge in the field, or working with these methods on astrophysical themes and data. Our IA@IRAP day aims to offer a space for interaction and feedback, and to create a dynamic around AI applications in astrophysics, measurements and instrumentation.

We will have two guest experts, Jonathan Sprauel (AI expert from IRT Saint-Exupéry, ANITI operational director) who will give a general introductory presentation on AI methods, approaches and limits, and Karin Dassas (Ecoinfo board member, CNRS, CESBIO) who will talk about the ecological impact of digital technologies in general.

A round-table discussion is planned for the day, which you are invited to attend.

Organising committee :

• Hui Yang, postdoctoral researcher , GAHEC, IRAP

• Rungployphan Kieokaew, AI spatial physics engineer, IRAP and Inria.

• Jihane Moultaka, astronome-adjoint, GAHEC, IRAP

09:00 → 09:05 Welcome

09:05 → 09:20 AI for SDU at OMP and the region

Orateur: Etienne Gondet (OMP)

Etienne_Gondet.pdf

09:20 → 10:30 Comment ça marche Chat GPT ? une introduction à l'Intelligence Artificielle

Tout ce que vous avez toujours voulu savoir sur l'Intelligence Artificielle, depuis l'optimisation et les modèles d'analyse d'image jusqu'au grands modèles de langage : quelles sont les différentes technologies et comment elles marchent, est-ce qu'on peut avoir confiance dans l'IA, quelles techniques d'IA je peux utiliser aujourd'hui dans mon métier, que font les chercheurs en IA, quels sont les impacts écologiques et sociétaux de l'IA.

Orateur: Jonathan Sprauel (IRT Saint-Exupéry)

Jonathan_Sprauel.pdf

10:30 → 10:50 AI or not AI ? Overview of methods developed at SISU

Orateur: Hervé Carfantan

Hervé_Carfantan.pdf

10:50 → 11:10 Coffee break

11:10 → 11:30 Physics models in Unrolling algorithms applied to Coded-Aperture Snapshot Spectral Imagers

Léo Paillet, Hervé Carfantan, Simon Lacroix, Antoine Monmayrant

We made a CASSI (Coded Aperture Snapshot Spectral Imager) optical simulator which allows to consider a realistic optical model. The model is then used to map input positions to output positions in order to adapt unrolling algorithms to realistic systems. Through this simulator and mapping function, we prove that the inverse problem of reconstructing hyperspectral scenes from CASSI acquisitions with unrolling algorithms is similarly solved for any type of CASSI system.

Orateur: Léo Paillet

Léo_Paillet.pdf

11:30 → 11:50 What can we learn from the XMM source catalog ?

The 4XMM catalogue contains more than 20 years of X-ray observations. However, exploiting this data on the catalogue scale is not straightforward since most of the automatically extracted sources do not have a precise classification. Our aim is to use machine learning, specifically a variational autoencoder, to embed the sources in a latent space that is more suitable for exploring and classifying the sources and finding exotic objects.

Orateur: Simon Dupourqué (GAHEC)

Simon_Dupourqué.pdf

11:50 → 12:10 Machine Learning-Led Classification of X-ray Sources Through Time-Domain Variability

Unusual patterns of variability in X-ray sources can hold clues to the identities of hard to classify sources, and can be indicators of rare and exotic astrophysical phenomena like magnetar outbursts and Quasi-Periodic Eruptions in AGN. The ability to adequately identify interesting sources through their time-domain variability can often be difficult with the large volumes of data being produced, and in the archives of large missions like XMM-Newton and Chandra. This problem will become ever more important with the anticipated launch of larger observatories like NewAthena in the next decade. Here I will present the preliminary results of an investigation into the use of classifiers for the identification of the rare phenomenon in AGN, Quasi-Periodic Eruptions and Quasi-Periodic Oscillations, and outline a route map for the development of an ensemble classifier aimed at providing robust predictions of sources solely through features extracted from their X-ray lightcurves.

Orateur: Robbie Webbe (GAHEC)

Robbie_Webbe.pdf

12:10 → 12:40 Unsupervised classification of galaxy spectra up to z~1.2. Towards galaxy spectral evolution ?

I will present our results of spectral galaxy classifications from GAMA and VIPERS surveys in the nearby universe and up to z~1.2 using an unsupervised method called Fisher-EM. This method is particularly well-suited to high-dimensional parameter spaces as is the case of galaxy spectra. Our approach enables us to highlight classes of galaxy spectra that allow to refine the bimodal distribution of galaxies and build an evolutionary tree up to z~1.2.

Orateur: Jihane Moultaka (GAHEC)

Jihane_Moultaka.pdf

12:40 → 14:00 Break: Lunch break (Buffet in Coriolis)

14:00 → 14:35 Panorama des activités IA sur ARIEL

La mission ARIEL de l'ESA vise à analyser la composition chimique des atmosphères exoplanétaires grâce à la spectroscopie de transit. Cette présentation se concentrera sur l'utilisation du Machine Learning pour transformer et améliorer les données ARIEL, en s'appuyant notamment sur des Data Challenges dédiés et des approches innovantes. Nous présenterons comment l'apprentissage automatique est utilisé pour traiter les différents niveaux de produit de la mission afin d'améliorer la résolution et de corriger les bruits. Nous aborderons les travaux actuels sur la modélisation chimique des atmosphères au travers des algorithmes IA utilisés. Enfin, nous parlerons des perspectives et des retours d'expérience sur l'utilisation du Machine Learning.

Orateur: Orphée Faucoz (CNES)

Orphée_Faucoz.pdf

14:35 → 14:55 RMI Analyst

We are developing a tool to classify images taken by ChemCam on Mars and rank their similarities.

Orateur: Olivier Gasnault (IRAP)

Olivier_Gasnault.pdf

14:55 → 15:15 Investigating the benefits of Deep Learning for particle pulse-shape discrimination in a space-borne silicon detector

The SP@M (for Solar Particles @ Mars) instrument onboard M-MATISSE will measure Solar Energetic Particles (SEPs, protons and electrons) throughout the Martian magnetosphere and atmosphere. While most of today's on-board instruments discriminate particles using coincidence processing between 2 or more silicon detectors arranged as a telescope, SP@M proposes to use a single thick silicon detector and to discriminate particle types (mainly electrons, protons and alphas ) using digital processing based on charge carrier collection time proportional to particle penetration length inside the detector. However, the first tests show that discriminating particles types can be challenging, particularly at low energies, as there is a confusion zone where electrons and protons have the same energies and collection times. We have investigated different Machine learning algorithms to take advantage of the difference of the pulse shapes between the particles types. We report their efficiency in discrimination and regression and we discuss the possibility of performing such a discrimination onboard a satellite.

Orateur: Pierre Devoto (IRAP/CNRS)

Pierre_Devoto.pdf

15:15 → 15:35 SPARTAI — AI-based forecasting pipeline for energetic electrions in the Earth’s radiation belts

Highly energetic electrons (at the MeV level) in the vicinity of the Earth’s radiation belts pose significant risks on satellites through single event effects and long-term radiation damage, in GEO, MEO, and LEO orbits. Accurate forecasting of electron fluxes in this environment is essential for risk mitigation and spacecraft operations. Over the past few decades, several physics-based models have been developed to forecast radiation belt conditions. Here, we develop an AI-based forecasting pipeline for MeV-level electron fluxes a few days in advance using NOAA’s space weather instruments and NASA-NOAA’s Geostationary Operational Environmental Satellites (GOES). We tested and benchmarked several architectures of machine learning and neural network architectures (e.g., CNN, LSTM, and Transformers), as well as several combinations of these to address data sparsity while optimising performance. Our preliminary results are rather promising, with an R2 score over 80 % for all L-shells (McIlwain L-parameter). Using data covering for at least one complete solar cycle, we will present our findings during extreme events. We will also consider European-based and the ESA’s space weather data for model training and for the development of the future operational forecasting pipeline. This work is a prototype product demonstration co-developed by Inria for Augura Space, a deep tech startup that delivers AI-powered space weather intelligence.

Orateur: Rungployphan Kieokaew (IRAP et Inria Centre de Paris)

Rungployphan_Kieokaew.pdf

15:35 → 16:00 Coffee break

16:00 → 16:45 What do we know about the environmental impact of research computing (including AI, but not limited to it)?

Cette présentation tentera d'apporter un éclairage sur l'impact environnemental que peuvent avoir les calculs faits à l'OMP, localement ou au sein de Calmip, Trex, ou les machines du GENCI. Elle abordera aussi l'impact de l'usage de l'IA générative pour aider à coder (IA4code), et la question des usages (IA4Green ?).

Orateur: Karin Dassas (CNRS, CESBIO)

Karin_Dassas.pdf

16:45 → 17:45 Round-table discussion