Strong gravitational lensing has become one of the main astrophysical probes of the extragalactic universe. It allows precise measurements of deflector galaxy masses. It acts as a cosmic telescope, revealing sources otherwise too dim or small to study. When lensed sources vary over time, strong lensing also offers a distance-ladder-independent measurement of the Hubble constant. However, these applications remain limited by the scarcity of known lenses and the absence of high-resolution imaging and spectroscopic redshifts for deflectors and sources. This limitation is now being overcome by the new generation of wide-field surveys. The Euclid Mission alone is expected to discover 100000 new lenses by the end of its six-year mission, ushering in an era of large-scale lens population studies. Yet current lens-finding techniques suffer from low purity and completeness, yielding hundreds of false positives per true lens and missing over half of lenses. And some lensing types, such as edge-on late-type galaxies (edge-on lenses), are entirely overlooked.
Therefore, to fully exploit these new surveys, we need to improve the purity and completeness of our lens-finding algorithms, and devise strategies to identify the more exotic lenses. That is the main goal of this thesis. We do this by performing lens searches on large data sets while devising and testing metrics that correctly evaluate our lens-finding techniques. We started by launching a massive search for edge-on lenses in particular, using 3600 deg² of r-band observations by the Ultraviolet Near Infrared Optical Northern Survey (UNIONS). We combined a convolutional neural network (CNN) trained on simulations of edge-on lenses, with expert visual inspection, finding 82 edge-on lens candidates and 50 non-edge-on candidates. We also estimated the prevalence of lensing in our parent sample via visual inspection of randomly selected galaxies (1/10000 and 1/30000 for edge-on lenses), allowing us to empirically evaluate our CNN model, and establishing a framework to compare lens-finding models.
Then, we led the exploitation of the Euclid Early Release Observation (ERO). By visually inspecting all the bright galaxies in the ERO Perseus cluster field, we found 16 high-quality lens candidates, and managed to model five of them. We extrapolated those five lenses to the entire survey and concluded that Euclid will discover between 70000 and 170000 new lenses, confirming the decade-old prediction that Euclid will find 100000 lenses. We also used an ensemble of CNN models to search for lenses in the remaining 16 ERO fields, finding 45 high-quality candidates in just 8.12 deg². We evaluated the lens search using the framework established in the UNIONS search, showing its potential as a systematic way of comparing models, which is fundamental when combining different model scores.
Beyond finding new lens candidates, this work contributed lens-finding methodologies, techniques, and tests that are now part of E
