2017 projects

Future space cosmological surveys will measure the weak gravitational lensing effect of large scale structures on the image of distant galaxies, providing new clues on the nature of dark matter, dark energy, and on the content of the Universe.

In practice, this requires the precise measurement of the shape of billions of faint and small galaxies and will requires “image denoising”. The work will consist in applying a denoising autoencoder (DA) to galaxy images with the objective of removing noise without altering the shapes of galaxies e.g. size and ellipticity.

An autoencoder is a specific type of Artificial Neural Network where the input and output layers are identical. The objective is to have the autoencoder learn how to build a new representation of the input data (the galaxy image) that has less noise while preserving the galaxy shape attributes.

The DA will be applied to simulated galaxy images with noise of different levels and types (Gaussian, Poisson). The alteration of shapes before and after applying the DA will be assessed using a galaxy shape measurement tool.


Il s’agit de mettre au point une méthode rapide (C + python) pour la détermination automatique des paramètres atmosphériques d’étoiles: température effective, gravité, vitesse de micro-turbulence, et métallicité.

Cette méthode s’appuyera sur des méthodes de minimisation robustes, inclura des prises de décision automatiques et des accès web à des bases de données extérieures.

Le code sera testé sur des données réelles et servira au plus grand survey actuellement en cours d’étoiles extrêmement déficientes en métaux dans le halo de notre galaxie pour comprendre les premières étapes dans la formation des galaxies.

Exemple de code existant


Accurately reproducing the observed stellar properties of dwarf spheroidal galaxies (dSphs), the smallest and faintest galaxies in our universe, represents a challenging task for the standard cosmological model.

The work proposed here will consist in understanding the impact on the stellar properties of dSphs (star formation history, metallicity distribution, metallicity gradient, stellar populations, rotation) of the interactions and collisions that dSphs may suffer during a Hubble time.

The study will consist in extracting dwarf galaxies formed in cosmological numerical simulations and make them collide, running idealised N-body simulations based on the chemo-dynamical-C/MPI-code GEAR (Revaz & Jablonka 2012, Revaz et al. 2016) which includes a complete treatment of the baryonic physics (gas cooling, star formation, chemical enrichment, supernova feedback).

The aim will be to explore the effect of parameters like the mass of the progenitors, the impact parameter and the relative velocity on the final properties of the dwarf galaxies. In this purpose, a set of python routines will be developed to allow for a comparison between the model predictions and the observations. The final goal will be to try to reproduce peculiar features observed in local group dwarf galaxies, like distinct stellar populations and metallicity gradients.


Strong lensing happens when the path of the light coming from a background galaxy or quasar is distorted by a foreground massive and compact object. The project is to search for strong lensing systems in the ~2 millions galaxy and quasar spectra of the BOSS and eBOSS catalogs. Spectra are provided with PCA templates of galaxy or quasar automatically fitted by the BOSS pipeline.

Strong lensing candidates are identified by looking for emission lines in the spectra not associated with the template, indicating the presence of an additional galaxy along the line of sight of the object. The goal is to conduct a systematic search for such emission lines in both galaxy and quasar samples and will use machine learning to obtain an automated classification of the lens candidates.



Quasars are supermassive black holes living at the centre of galaxies. The Dark Energy Spectroscopic Instrument (DESI) will obtain 700’000 quasar spectra to map the Universe at redshift z>2.1 using a recent technique called the Lyman-alpha forest.

The detection of atypical spectra including broad absorption line and damped lyman-alpha systems will be essential for cosmological analyses. The project is to exploit the large quasar spectra database from BOSS and eBOSS (~100’000 object using machine learning algorithm to setup an automatic finder for atypical quasars to be used for DESI.



Il s’agit de tester et d’améliorer un algorithme d’identification de filaments, c’est à dire de galaxies corrélées spatialement en 3 dimensions, dans des champs dans lesquels nous avons les redshifts des systèmes. La méthode proposée s’appuie sur une technique de tesselation. Elle sera testée sur des simulations numériques et des données observées.

Code: http://www2.iap.fr/users/sousbie/web/html/indexd41d.html