Get started
Work

Nouvelles techniques d’apprentissage automatique pour répondre au volume et à la complexité des données du relevé Rubin LSST : séparation d’images de galaxies superposées et classification dans les flux d’alertes relatives aux événements transitoires

Year: 2024
Type: dissertation
Abstract: Designing novel algorithms to handle the massive volumes of complex data from the next generation of cosmological surveys, is the need of the hour. In particular, Machine Learning can address some of the most challenging problems we are expected to face. In the context of the Legacy Survey of Space and Time (LSST) at the Vera C. Rubin Observatory, this work presents a couple of novel machine learning-based algorithms for the following: Firstly, with the increase in object density due to the depth of the survey, galaxies will tend to overlap in our images - an effect called blending. I present an algorithm called MADNESS, developed with the LSST-Dark Energy Science Collaboration (DESC), to solve the inverse problem of deblending where we extract the isolated galaxies from a blended scene. MADNESS uses deep learning to extract data-driven priors and achieves a better performance than commonly used algorithms, in terms of flux, morphologies, shapes, and colors of galaxies. Secondly, due to the wide-field nature of the survey, LSST will scan large areas each night and detect transient events by looking for fluctuations in brightness in the sky. LSST will detect a massive volume of up to 10 million such ‘alerts’ per night, which requires intermediary software called brokers to identify interesting events. In the context of the Fink broker, I present a kilonova detection module that looks for fast transients and can be used to detect multi-messenger candidates with a corresponding gravitational wave detection.
Language: English
Open Access status: green
OpenAlex ID: W7080190617
Cites:
Cited by:
Related to: