During the last decade, cosmology has entered a precision era, leading to the prevalence of the standard cosmological model, ΛCDM. Nevertheless, the main ingredient of this model, dark energy, while dominating the energy budget of the Universe, remains mysterious and its comprehension is the current Graal of the domain. The next generation of cosmological surveys, among which Large Synoptic Survey Telescope (LSST) & Euclid, both starting in 2022, are a major step toward in observational cosmology toward the parametrization of dark energy. We would like to offer to a PhD student the possibility to contribute to the development of state-of-the-art cosmological analyses within the APC LSST group in the framework of both the LSST Dark Energy Science Collaboration (DESC) and the AstroDeep ANR project.

The scope of this work is related to weak gravitational lensing, the deflection of light from distant sources due to the bending of space-time by matter along the line of sight. Weak lensing is one of the main probes used to study dark energy (with strong lensing, supernovae Ia, galaxy clusters, and large scale structures studies). Its detection and quantification as a function of distance to us will enable a measurement of the growth rate of cosmic structures, and therefore higher constraints on the nature and evolution of dark energy. For such purpose, the unprecedented volume of multi-band data from LSST combined with the high resolution and near IR imaging of Euclid will be perfect.

However this increase in sensitivity compared to previous surveys will also bring its share of new difficulties. As more and more galaxies and stars populate the images, the local density increases and the projected objects naturally overlap, a challenging phenomenon referred to as blending. Blending will affect more than 60% of the galaxies in LSST and needs to be addressed for proper weak leasing measurements. Our group has developed a novel approach, based on deep learning, that uses the multi-band images of LSST and Euclid fed to a Bayesian neural network to separate overlapping galaxies before measuring their shape. This work has demonstrated the feasibility of the approach in a configuration of two simulated galaxies, and the goal of the internship offered during summer 2020 would concern the application of the existing tool to real data from available surveys ; and one goal of the following thesis would be to extend the approach to more realistic configurations and scientific pipelines.

Going beyond the first task of separating the galaxies to measuring their shape, we would like to attempt a direct, statistical, measurement of the average local shape of blended objects. The problem can be addressed in a number of ways, but one promising direction would be to investigate is the use of Bayesian deep nets (BNN) on this very problem as well. In the scope of our starting AstroDeep ANR project, collaborations with statisticians and computer scientists on the promising aspects of uncertainties quantifications within BNNs will be facilitated. The second aspect of the thesis work would thus be to build on the framework of an ongoing PhD thesis and extend it on the characterization of this approach for weak lensing analyses and its integration within the DESC collaboration.

LSST and Euclid data will become available for science in 2022, during the PhD thesis (LSST is starting its commissioning in 2021), making these studies all the more interesting as the scientific environment is becoming increasingly dynamic. As the data will be gathered, one of the foreseeable goals of the work we advocate for in the AstroDeep project and for this thesis work is to attempt the use of a « multi- instrument, multi-wavelength » survey: shapes, colors could be used in order to greatly improve our precision on shear measurements. If successful, the implications of this work on a joint, multi-survey, multi-wavelength analysis could allow to reduce drastically the bias on shear measurements and release to the community an essential tool for weak lensing studies.