• 3-month internship possibly followed by a 3-year PhD thesis.
  • start: Spring 2021
  • deadline: February 2021
  • contacts: Cécile Roucelle, Eric Aubourg, Alexandre Boucaud

Dark energy studies with the Vera Rubin Observatory LSST & Euclid - Developing a combined cosmic shear analysis with Bayesian neural networks.

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, remains mysterious while dominating the energy budget of the Universe. Its comprehension is the current Graal of this domain. The next generation of cosmological surveys, among which Legacy Survey of Space and Time (LSST) of the Vera Rubin Observatory (on the ground) & Euclid (in space), both starting in 2022, are in that regard the most important projects for the next 10 years. These surveys, when combined, will map thousands of square degrees of sky in a multiwavelength manner with sub-arcsec resolution. This will result in the detection of several tens of billions of sources, enabling a wide range of astrophysical investigations and providing unprecedented constraints on the nature of dark energy and dark matter. The scope of the PhD topic is at the crossing of the two surveys. More precisely, the PhD topic discussed here (and the preceding internship) is focused on developing analyses for weak gravitational lensing combining the data of LSST and Euclid.

The gravitational lensing corresponds to the deflection of light from distant sources (background galaxies) due to the bending of space-time by matter along the line of sight, resulting in distortions and displacements of their image. The statistical study of weak gravitational lensing distortions at large scales provides a “mapping” of the matter (dark or visible) between the observer and source (more accurately, the effect of coherent deformation described here is called cosmic shear). This type of measurement gives a window on the properties and the evolution of cosmic structures as well as the geometry of the Universe. Its study can therefore bring higher constraints on the origin of the current accelerated expansion of the Universe that led to the notion of dark energy. In the absence of systematic errors, weak lensing is even recognised as the single most constraining probe of dark energy. As such, it is one of the main science drivers for both LSST and Euclid.

However with an increased sensitivity compared to previous surveys, LSST and Euclid will bring their share of new challenges to allow the proper use of weak gravitational lensing. For example, the sheer volume of produced data requires the development of new types of analyses that allow a more efficient processing. Plus novel complexities arise. To address these challenges, our group has started to develop an approach based on Bayesian neural networks (BNNs, [1]). The BNNs offer a formalism to quantify and propagate uncertainties associated with deep neural networks models and also with the data itself, which are both key for cosmological analyses.

To give an example, as more and more galaxies and stars populate the images, the local density increases and the projected objects naturally overlap. This phenomenon, referred to as blending, impedes the ability to measure properly the shapes of individual objects and will affect more than 60% of the galaxies in LSST. To address this issue, deep learning brings a possible solution, with an efficient use of the joint multi-band processing of LSST and Euclid images. The images are fed to a BNN to separate overlapping galaxies before measuring their shape and this approach brings an improvement to the use of one of the surveys alone (as it brings the complexity but also the power of a multi-resolution and multiwavelength approach to the problem). We have demonstrated the feasibility of this approach in a configuration with several simulated galaxies per image ([2]). This approach is still further tested on more realistic configurations. Depending on the start date for the pre-thesis internship, the main topic to pursue would be the quality assessment of the deblending procedure on realistic galaxy fields simulations, with a focus on the LSST-Euclid combination.

Going beyond the first task of detecting and separating individual galaxies, the successful candidate would work on a direct statistical measurement of the average shape of blended objects with BNNs, leading to a local estimate of the cosmic shear. This would be the main topic of the PhD, going from the development of the image processing pipeline itself to the comparison and exploitation of these algorithms on LSST and Euclid simulations and real data once available.

From a local environment perspective, this PhD will occur in the scope of the AstroDeep project started in October 2019 at APC. This project intends to build a collaboration between cosmologists, statisticians and computer scientists, focused on the quantification of uncertainties in cosmological analyses with Bayesian neural networks. The student joining our group would beneficiate from such collaboration to build knowledge in an emergent field, working on the comprehension, characterization and use of the BNNs, and also apply these exciting techniques to astrophysical images to perform a multi-instrument and multi-wavelength joint analysis of galaxy fields. If successful, the implications of this work could drastically reduce the bias on cosmic shear measurements and release an essential tool for observational cosmology to the community. Not to mention that 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 will be extremely dynamic.

[1] Tom Charnock, Laurence Perreault-Levasseur, François Lanusse, Bayesian Neural Networks – arXiv:2006.01490
[2] Bastien Arcelin, Cyrille Doux, Eric Aubourg, Cécile Roucelle, Deblending galaxies with Variational Autoencoders: a joint multi-band, multi-instrument approach, MNRAS, staa3062 – arXiv:2005.12039v1