Teaching machines how to find strongly lensed galaxies in cosmological sky surveys.

Analyzing gravitationally lensed objects enables a wide range of physical and cosmological applications, such as probing the dark matter content in galaxies and clusters or measuring the Universe's expansion rate.The precision of these applications can be improved drastically with a larger number of lenses. For this reason, finding new strong gravitational lensing systems is crucial. Traditionally, gravitational lenses were found with meticulous visual inspections of imaging datasets. However, the next generation of large-scale imaging surveys will produce such a large amount of data that visual inspection will be unpractical, thus motivating the development of efficient automated detection methods to handle large datasets and improve the accuracy of the different applications.The main topic of this thesis is to develop and improve our lens finding algorithms and to study new strong gravitational lenses in large-scale imaging surveys.The first part of this thesis has been to develop a new tool that allows the production of lens simulations. Deep learning algorithms, particularly convolutional neural networks (CNNs), have recently proven their efficiency in detecting lensing systems. In particular, CNNs can be adapted to find a large variety of lenses at once, making them especially suited to be part of automated detection pipelines. However, convolutional neural networks usually require large sets of images to be trained. Unfortunately, the number of known lenses is at this time too low to constitute a sufficient training set. For this reason, robust and flexible tools to simulate realistic lenses have to be developed to generate training sets. Lens simulations must, however, be as realistic as possible to avoid biases.In this thesis, we propose a simulation tool that enables the production of large sets of realistic lens simulations. This tool has been designed to be flexible and, thus, enables the production of lens simulations with different types of deflectors for any imaging survey.The second part consists in separating the light of a lensed source from the foreground object. Indeed, in some cases, the lens features are hidden by the light of the foreground object and deblending might help to identify the lenses help identify lenses with small image separation and enables photometric redshift measurements or the initialization of lens models. This thesis presents a new data-driven method for deblending strong gravitational lenses based on neural networks.Finally, due to the low occurrence rate of strong lensing, the false positive rate of detection algorithms is still a significant challenge. Therefore, we present different tools that enable the inspection of large sets of candidate images. In addition, evaluating the quality of lens candidates is somewhat subjective since the features that define a lens may differ for different experts. Therefore, we propose a set of grading guidelines that can be used for the subsequent ground-based imaging lens searches in this thesis.The CNN-based classifier, the deblending algorithm, and the simulation and visualization tools are part of an automated lens-finding pipeline. This flexible pipeline can be adapted to future large-scale surveys. We discuss in this thesis its first applications to the Canada-France Imaging Survey and the Dark Energy Survey, which led to the discovery of 133 and 403 high-quality lens candidates, respectively.