Machine learning for strong lensing: paving the way for the Euclid mission

In this Thesis, we explore the possibilities and limitations of the application of machine learning techniques to the analysis of strong gravitational lensing events by galaxies and galaxy clusters. This work is done in preparation of the survey that the Euclid mission is conducting, and that will cover one third of the sky in six years, delivering the images of more than one billion galaxies. Given the large volume of data, the identification of strong lensing events and their subsequent modelling will be necessarily based on automated methods. The main objective of this work is characterizing the performance of these techniques and exploring new possible applications. In the first part of the Thesis, we investigate the impact of training sets on the ability of convolutional neural networks to identify lens candidates. We evaluate this using Euclid mock observations, by training three architectures on simulated data, and on the Early Release Observations of the Perseus cluster, that are among the first publicly images released by the Euclid collaboration. In the second part of the Thesis, we focus on strong lensing by galaxy clusters. We develop a new version of the code SkyLens, for simulating cluster-lenses, and we validate the procedure for generating the clusters' properties by comparing the lensing properties of a simulated cluster to those of a real cluster of similar mass distribution and redshift. Moreover, these simulations are useful for the validation of automated classification and modelling methods, because we generate a catalog of properties of cluster members and lensed sources along with the images. Finally, we present a novel method, currently under development, that relies on equivariant graph neural networks for identifying the multiple images of lensed background sources. This work is important in the context of characterizing machine learning models before their widespread application to Euclid data.