Personal web page : http://jstarck.cosmostat.org
Laboratory link : http://www.cosmostat.org
The Euclid satellite, to be launched in 2022, will observe the sky in the optical and infrared, and will be able to map large scale structures and weak lensing distortions out to high redshifts. Weak gravitational lensing is thought to be one of the most promising tools of cosmology to constrain models. Weak lensing probes the evolution of dark-matter structures and can help distinguish between dark energy and models of modified gravity. Thanks to the shear measurements, we will be able to reconstruct a dark matter mass map of 15000 square degrees. Mass mapping entails the construction of two-dimensional maps using galaxy shape measurements, which represent the integrated total matter density along the line of sight. Small- field mass maps have been frequently used to study the structure and mass distribution of galaxy clusters, whereas wide-field maps have only more recently become possible given the broad observing strategies of surveys like CFHTLenS, HSC, DES, and KiDS. Mass maps contain significant non-Gaussian cosmological information and can be used to identify massive clusters as well as to cross-correlate the lensing signal with foreground structures.
A standard method to derive mass maps from weak-lensing observations is an inversion technique formulated by Kaiser & Squires [2]. It has many limitations, however, including the need to smooth the data before (and often after) inversion, thereby losing small-scale information. An alternative method called GLIMPSE has been developed in the CosmoStat laboratory based on sparse reconstruction that avoids this problem and improves the recovery of non-Gaussian features [3, 4]. The algorithm has been tested on simulations and was also recently used to study the A520 merging galaxy cluster with Hubble Space Telescope data [5]. More recently, machine learning has emerged as a promising technique for mass map recovery [6].
The goal of this thesis is to i) compare this technique to the state of the art and investigate if it can be used in practice, ii) extend the method for spherical data, and iii) develop a new machine learning approach to estimate the cosmological parameters. At the core of this new statistical framework will be the development of fast and differentiable cosmological simulations capable of emulating the Euclid survey under various cosmologies. This simulation tool will be based on the FastPM N-body simulation code [7] and implemented directly in the TensorFlow machine learning framework, yielding a differentiable physical forward simulation pipeline which can be directly interfaced with deep learning components or with inference techniques relying on having access to the derivatives of the simulation.
As part of the CosmoStat Laboratory, located at CEA Saclay, the successful candidate will be embedded in a leading French research group, heavily involved in the preparation of the Euclid space mission, and with a long tradition of developing cutting-edge statistical tools for the analysis of astronomical and cosmological data.
1. Bartelmann, M. & Schneider, P. 2001, Phys. Rep., 340, 291. ?
2. Kaiser, N. & Squires, G. 1993, ApJ, 404, 441. ?
3. Leonard, A., Lanusse, F., & Starck, J.-L. 2014, MNRAS, 440, 1281.
4. Lanusse, F., Starck, J.-L., Leonard, A., & Pires, S. 2016, A&A, 591, A2.
?5. Peel, A., Lanusse, F., & Starck, J.-L. 2017, ApJ, 847, 23.
6. Niall Jeffrey et al, submitted. https://arxiv.org/abs/1908.00543
7. Y. Feng, M. Yat Chu, U. Seljak, and P. McDonald. MNRAS, 463(3):2273–2286, 2016.