Introduction

• Jean-Luc Starck (CosmoStat, CEA Paris-Saclay)

Room: Room Stelios Orphanoudakis

Radio weak-lensing shear measurements using deep learning

• Priyam Tripathi

Room: Room Stelios Orphanoudakis The limited shape measurement methods in the radio waveband are either computationally intensive or fail to achieve the accuracy required for future surveys. In this talk, I will present a supervised deep learning framework, dubbed DeepShape, that measures galaxy shapes with the necessary precision while minimizing computational expenses. DeepShape is made of two modules. The first module is a plug-and-play (PnP) image reconstruction algorithm based on the half-quadratic splitting method (HQS), dubbed HQS-PnP, which reconstructs images of isolated radio galaxies. The second module is a measurement network that predicts galaxy shapes using the point spread function and reconstructed image pairs. The HQS-PnP algorithm outperforms the standard multiscale CLEAN algorithm across several tested metrics, particularly at low PSNR values. In terms of shape measurement, DeepShape recovers galaxy shapes with an accuracy comparable to the leading radio shape measurement method, RadioLensfit, while significantly reducing the prediction time from ∼4 minutes to ∼220 milliseconds. I will also show some additional preliminary results on how the methodology could be extended to the case of non-isolated galaxies and how to deal with challenges such as far-sidelobe confusion noise and galaxy deblending.

Generative model for shear inference

• Ezequiel Centofanti

Room: Room Stelios Orphanoudakis Next-generation radio interferometers, such as the SKA, will observe the radio sky with unprecedented sensitivity and resolution. Their wide sky coverage will also enable weak lensing studies using radio data. Radio weak lensing not only complements optical observations, but also allows access to higher redshifts. However, traditional shear estimation methods—based on measuring the ellipticity of observed galaxies—are not directly applicable to radio data. Interferometric observations are acquired in the Fourier domain, where galaxy images are delocalised, making shape measurements non-trivial. Moreover, even in the image domain, radio galaxy morphologies differ significantly from their optical counterparts, making parametric shape fitting unreliable and prone to model bias. To address these challenges, we propose a cosmic shear inference method based on deep generative models, which avoids parametric fitting and explicit shape measurements. In this presentation, I will outline the main challenges involved in building this inference pipeline and share preliminary results.

A Decentralized Framework for Radio-Interferometric Image Reconstruction

• Sunrise Wang

Room: Room Stelios Orphanoudakis A Decentralized Framework for Radio-Interferometric Image Reconstruction Abstract The advent of large aperture arrays, such as the ones currently under construction for SKAO, allows for observing the Universe in the radio-spectrum at unprecedented resolution and sensitivity. To process the enormous amounts of data produced by these telescopes, scalable software pipelines are required. In this presentation, I will introduce a framework that allows for decentralized radio-interferometric image reconstruction, parallelizing by spatial frequency. This is achieved by creating pseudo full resolution problems for each node by using the local visibilities together with previous major cycle reconstructed images from the other nodes. I will show results when applying this proposed framework to both multiscale CLEAN and sparsity regularized convex reconstruction in the context of two partitions, for which it has shown to perform well. I will also show some additional preliminary results when scaling this method up to a larger number of partitions, discuss some of the current roadblocks, and also discuss some additional applications of separating the reconstruction by spatial frequency.

Plug-and-play weak lensing mass mapping with fast uncertainty quantification

• Hubert Leterme (ENSICAEN)

Room: Room Stelios Orphanoudakis In this talk, I will present a plug-and-play (PnP) approach for estimating the dark matter distribution from weak gravitational lensing data, using noisy shear measurements. Our method is designed to provide accurate and efficient mass maps without the need to retrain deep learning models for each new galaxy survey or sky region. Instead, a single model is trained on simulated mass maps corrupted by Gaussian white noise. We show that a well-chosen data fidelity term accelerates convergence to the algorithm's fixed point. Additionally, we adapt a fast uncertainty quantification (UQ) method, based on order-2 moment networks, to the PnP framework. Unlike existing UQ approaches in this context, this method does not rely on posterior sampling, which is often computationally intensive. We benchmark our method against both model-driven and data-driven mass mapping techniques, and show that it achieves state-of-the-art reconstruction accuracy while producing smaller error bars, all with increased flexibility.

Discussion Room: Room Stelios Orphanoudakis

Mass mapping and cosmological inference with higher-order statistics

• Andreas Tersenov

Room: Room Stelios Orphanoudakis In this talk, I will present ongoing work toward improving cosmological inference from weak lensing through a synergy of forward modeling, simulation-based inference (SBI), and higher-order statistics. First, I will discuss the impact of mass mapping algorithms on the accuracy and robustness of cosmological parameter estimation. I will then introduce a framework that combines SBI with the BNT (backward normalizing transform) to directly infer cosmological parameters from noisy shear fields, allowing for flexible, likelihood-free inference while capturing non-Gaussian features, and show the impact of unmodelled baryonic effects on the inferred parameters. Finally, I will highlight the role of theory-informed higher-order statistics, specifically the wavelet l1-norm, and their validation against suites of cosmological simulations, on the level of cosmological contours. These tools together form a robust framework for extracting maximal cosmological information from current and next-generation weak lensing surveys, including Euclid and UNIONS.

Wavelet l1norm: A Theory-Driven Approach Beyond Two-Point Inference

• Vilasini Tinnaneri Sreekanth

Room: Room Stelios Orphanoudakis Weak gravitational lensing is a key cosmological probe of the large-scale matter distribution, yet conventional two-point statistics are insufficient to capture the non-Gaussian features imprinted by nonlinear structure formation. This thesis develops a novel framework based on the wavelet ℓ1-norm to extract higher-order information from weak lensing convergence maps. Leveraging predictions from Large-Deviation Theory, we construct an analytical model for the one-point PDF of wavelet coefficients and use it to generate non-Gaussian maps consistent with theoretical expectations, bypassing the need for N-body simulations. The method enforces consistency with the power spectrum and is designed to incorporate inter-scale correlations and observational systematics such as baryonic effects and intrinsic alignments. Ongoing work focuses on validating the theoretical predictions through comparisons with simulation-based constraints. This framework offers a theory-driven, flexible approach for extracting maximal cosmological information from current and future weak lensing surveys.

Deep Denoising and Signal Restoration for Astrophysical Spectral Cubes: From Simulations to Real Data

• Arnab Lahiri (FORTH)

Room: Room Stelios Orphanoudakis Spectral cube data obtained from instruments such as ALMA are fundamental for understanding the structures and dynamics of astrophysical objects. However, noise, beam convolutions, and other artifacts can mask important astrophysical features. This is more evident in high-redshift observations as large cosmological distances, high levels of noise and lower resolution beams may hinder signal extraction for scientific analysis, This work compares and optimizes denoising algorithms for spectral cube data, including Principal Component Analysis, blind source separation methods like Independant Component Analysis, iterative 2D-1D wavelet transforms, and U-Net based deep learning. These are being benchmarked to establish the performance of the methods for the improvement of the signal-to-noise, flux conservation, and noise reduction while preserving spatial and spectral features on synthetic, simulated, and real data. Methods. We apply the denoising techniques to the toy spectral data of rotating galaxies, mock IFU cubes from FIRE simulations with spectral lines characteristic of ALMA, and ALMA observations of W2246-0526.

Unsupervised Cloud Removal and Change Detection on Multi-Temporal Images via Unrolled Tensor Decomposition Network

• Anasthasia Aidini

Room: Room Stelios Orphanoudakis The increasing availability of multi-temporal remote sensing data offers unprecedented opportunities in monitoring environmental changes induced by natural disasters. However, a common issue in satellite imagery is cloud cover and cloud shadows, which can hinder further analysis. Given the high-dimensionality of satellite image time series and the lack of cloud-free observations and ground truth labels in real-world scenarios, we propose an unsupervised tensor-based unrolled network that simultaneously reconstructs cloud-occluded regions and detects the effects of extreme events. By leveraging algorithmic unrolling to benefit from both tensor analysis and deep learning techniques, we impute the missing values via a low-rank tensor decomposition approach while learning a feature space representation. Comparing the images before and after the event in the learned space enables effective change detection. Experiments on real multi-temporal Sentinel-2 images demonstrate the effectiveness of the proposed method in simultaneously removing clouds and detecting changes caused by fires.

Antenna Array Optimization for Radio Interferometry: Towards a Machine Learning-Guided Approach

• Manal BENSAHLI (CosmoStat, CEA Paris-Saclay)

Room: Room Stelios Orphanoudakis The quality of images produced by a radio interferometer strongly depends on the geometric layout of the individual antennas. Enhancing the UV-plane coverage through an optimal configuration improves the synthesized beam shape and enables a more accurate reconstruction of the sky, with reduced artefacts. In this work, we investigate several strategies for optimizing antenna placements using simulated observations of synthetic sky models. We rely on heuristic methods such as Particle Swarm Optimization (PSO) to minimize physically relevant beam metrics, including the full width at half maximum (FWHM) and the side-lobe level (SLL). Each candidate configuration is assessed through a complete imaging pipeline: baseline calculation, time- and frequency-dependent UV coverage, beam synthesis, and quantitative beam evaluation.In parallel, we initiate the development of a machine learning approach that aims to predict beam quality directly from antenna distribution parameters. This approach is intended to accelerate the configuration search process, with the longer-term goal of enabling efficient, ML-guided design of optimized interferometric arrays.

Improving Weak Gravitational Lensing Using Kinematic Information from Galaxies

• Jordy Ram

Room: Room Stelios Orphanoudakis In this talk, we introduce the concept of kinematic lensing to infer the intrinsic galaxy shapes and their deformation due to the matter distribution in the line of sight. Traditional weak lensing, in which the intrinsic galaxy shape is unknown, uses a significant number of galaxies in order to make an estimate of the galaxy shear. However, the shape noise will remain a source of variance in the measurements. Additionally, astrophysical systematics like intrinsic alignment, in which galaxies are aligned due to the tidal field, can provide a bias. Kinematic lensing is expected to reduce the variance from shape noise and the sensitivity to biases from intrinsic alignment significantly. We explore an existing kinematic lensing method called MIRoRS, which uses the symmetry in the velocity field of disk galaxies to determine the shear in the galaxy image. We have tested MIRoRS, along with an improved method, on mock galaxies from MaNGIA, a mock MaNGA sample generated using the Illustris TNG50 simulation. Furthermore, we investigated whether this kinematic lensing method could also be extended to elliptical galaxies, considering their relatively symmetric velocity fields. The preliminary results indicate that comparable shear constraints can be achieved with disk and elliptical galaxies, potentially increasing the number of galaxies that can be used with this kinematic lensing method. However, more realistic data is required to confirm this.

Discussions Room: Room Stelios Orphanoudakis

Inertial Non-Convex Optimization Algorithms Meet Neural Network-Based Inverse Problems

• Jalal Fadili (ENSICAEN)

Room: Room Stelios Orphanoudakis In this talk, I will focus on non-convex minimization problems via inertial second-order (in-time) dynamics and how they can prove valuable when solving inverse problems with neural network-based methods. I will first discuss several theoretical and practical issues fo these algorithms, including convergence, convergence rates and trap avoidance properties. I will then turn to discussing how to bridg the worlds of optimization and that of inverse problems to provide convergence and recovery guarantees for a class of neural network-based methods (DeepInvese). I will also a precise characterization of the network architecture to benefit from these guarantees. This provides a first step towards the theoretical understanding of the interplay between the optimization dynamics and neural networks in the inverse problem setting.

Detection of Fast Radio Bursts using hybrid neural architecture

• Sara El Bouch (OCA)

Room: Room Stelios Orphanoudakis FRBs-millisecond-duration radio pulses from distant galaxies-offer unique insights but remain challenging to detect due to overwhelming radio frequency interference. Traditional detection methods and conventional machine learning techniques struggle with extreme class imbalance and data volume challenges. Our solution treats FRB detection as an open-set anomaly detection problem, leveraging a dual-headed neural network that combines the strengths of both generative and contrastive learning paradigms. In this talk, I will present an approach to detecting Fast Radio Bursts (FRBs) using a hybrid contrastive-generative neural architecture.

UNIONS: First Cosmic Shear Constraints from 3000 deg² of Northern Sky

• Cail Daley (CEA)

Room: Room Stelios Orphanoudakis The Ultraviolet Near-Infrared Optical Northern Survey (UNIONS) is a multi-band optical survey that will cover over 4800 deg² of northern sky. Combining data from three wide-field Hawaiian telescopes (CFHT, Pan-STARRS, and Subaru), UNIONS' sky coverage and excellent image quality makes it an ideal dataset for weak lensing studies, providing cosmological constraints that are both independent of and complementary to those from Southern Hemisphere like DES and KiDS. In this talk we will present the first cosmic shear constraints from a 3000 deg² non-tomographic catalog containing 100 million galaxies, with which we expect to obtain a 6% measurement on the structure growth parameter S₈. We will discuss calibration tests of the PSF, galaxy shapes, and the redshift distribution, as well as prospects for future tomographic analyses that will provide even tighter constraints and bridge the gap between Stage III surveys and Euclid.

Overview of the ESA Euclid mission in the context of ESA Science today

• Valeria Pettorino (ESA)

Room: Room Stelios Orphanoudakis Abstract : After a brief introduction of the ESA Science programme I will focus on an overview of the ESA Euclid mission: I will describe mission objectives, challenges, main milestones so far, the first data release and next steps. I will also mention a few ESA programmes dedicated to interns, graduate students, and postdocs. bio: I am a physicist, and I work as Euclid Project Scientist at the European Space Agency (ESA), in ESTEC, Netherlands. After my PhD (Naples, Italy, 2005), I worked in Torino, Trieste (Italy), Heidelberg (Germany), New York (USA), Geneva (Switzerland). I joined CEA Paris-Saclay (France) in 2016 in the CosmoStat Lab, and then ESA in 2023. I have worked for ESA Planck space mission (2009-2018) and for ESA Euclid since 2007. My main expertise is in cosmology and dark energy, testing theories with observations.