Thèse Modèles de Fondation et Apprentissage Profond pour la Reconstruction Tomographique Jointe des Structures Sombres avec les Données du Télescope Spatial Euclid H/F - Université Paris-Saclay GS Physique

Établissement : Université Paris-Saclay GS Physique
École doctorale : Astronomie et Astrophysique d'Ile de France
Laboratoire de recherche : Astrophysique Instrumentation Modélisation
Direction de la thèse : Jean-Luc STARCK ORCID 0000000321777794
Début de la thèse : 2026-10-01
Date limite de candidature : 2026-04-30T23:59:59

Ce projet de thèse a pour objectif de développer une nouvelle génération de modèles de fondation physiquement informés et de méthodes d'apprentissage profond dédiés à la reconstruction tomographique jointe des structures de matière noire issues des données d'Euclid. En apprenant des représentations multi-échelles et multi-redshifts à partir de simulations réalistes, ces modèles intégreront les effets instrumentaux, les a priori physiques ainsi que la quantification des incertitudes dans un cadre méthodologique unifié.
Ces travaux permettront d'améliorer significativement la récupération des structures cosmiques non gaussiennes et d'évaluer précisément l'impact de ces reconstructions avancées sur l'inférence des paramètres cosmologiques. À terme, ce projet contribuera directement à l'exploitation scientifique de la mission Euclid et à la compréhension des origines des grandes structures de l'Univers.

Understanding the origin and evolution of cosmic structures is one of the central challenges of modern astrophysics and cosmology. The three-dimensional distribution of dark matter encodes fundamental information on the nature of gravity, the properties of dark energy, and the initial conditions of the Universe. Weak gravitational lensing, by measuring the coherent distortions of distant galaxies, provides a unique and direct probe of the dark matter distribution across cosmic time.
The ESA Euclid space mission is designed to map the large-scale structure of the Universe with unprecedented precision, enabling tomographic reconstructions of dark matter over a large fraction of the sky. However, exploiting the full scientific potential of Euclid requires overcoming major methodological challenges. Tomographic mass reconstruction is a severely ill-posed inverse problem, affected by instrumental noise, survey geometry, spherical projection effects, redshift uncertainties, and complex non-linear couplings between redshift slices. In addition, standard reconstruction pipelines rely largely on Gaussian assumptions and two-point statistics, which fail to capture the rich non-Gaussian information contained in cosmic structures such as clusters, filaments, and voids.
Recent advances in deep learning and foundation models open new perspectives to address these limitations. By learning multi-scale, multi-redshift representations from large ensembles of realistic simulations, such models can approach the reconstruction problem in a global and joint manner, integrating instrumental effects, cosmological signal properties, and uncertainty quantification within a unified framework.

The central objective of this PhD is to develop physically informed foundation models for the joint tomographic reconstruction of dark matter structures from Euclid weak lensing data, enabling a more faithful recovery of non-linear structures and a robust propagation of uncertainties, ultimately improving cosmological inference.

The PhD will go beyond task-specific neural networks and classical mass-mapping approaches by developing foundation models designed to:
jointly reconstruct 2D and 3D dark matter maps across redshift,
incorporate spherical geometry and survey masks,
integrate physically motivated priors and instrument models,
provide calibrated, spatially resolved uncertainty estimates.
The approach will build on recent advances in generative modeling (e.g. diffusion models), algorithm unrolling, and self-supervised learning, while maintaining strong interpretability and physical consistency. The models will be trained and validated on realistic Euclid simulations, with a clear path toward application to real mission data.

Research plan and Timeline:
Year 1 - Foundations and Problem Formalization
- Review of weak lensing theory, tomographic reconstruction methods, and Euclid instrumental effects
- Study of existing inverse-problem and deep learning approaches in cosmology
- Mathematical and statistical formalization of the joint tomographic reconstruction problem
- Familiarization with Euclid simulations and data infrastructure
Year 2 - Development of Foundation Models
- Design of multi-scale, multi-redshift neural architectures adapted to spherical data
- Integration of physical priors and instrumental models into learning frameworks
- Development of generative and hybrid physics-AI approaches
- Initial training on simulations and performance benchmarking
Year 3 - Validation, Cosmological Applications, and Valorisation
- Bias analysis, robustness tests, and uncertainty calibration
- Application to Euclid-like or precursor datasets
- Extraction of non-Gaussian statistics and assessment of cosmological impact
- Publications, open-source code release, and conference presentations

Un diplôme d'ingénieur ou M2 recherche en astrophysique, traitement du signal, ou apprentissage automatique est nécessaire. Être confortable avec le développement logiciel (au moins en Python) et les outils de développement ouverts et collaboratifs sera un atout important (e.g. GitHub).