Résumé / Abstract Journal-club_Doctorants

Journal-club Doctorants / Journal-club PhD students

« Galaxy evolution modelling with simulated images, Bayesian inference and dimensionality reduction through neural network. »

Florian Livet
Institut d'Astrophysique de Paris (Paris, France)

The study of galaxy evolution has shown the existence of a great variety of morphologies: ellipticals, lenticulars, spirals with or without bar and irregulars. We are seeking to understand how each morphological type has evolved and what are the possible transformations between the different populations. Inverse modelling is the common approach to study galaxy evolution: catalogs of fluxes and sizes of galaxies obtained from deep field surveys (e.g. CFHTLS) are used to understand the evolution history of galaxy population. Unfortunately, the observed galaxies are subject to many selection biases: cosmological dimming (1+z)^-4, Malmquist, Eddington, expansion of the Universe, dust extinction, occultation and confusion. The main difficulty of this approach is to model these highly correlated biases and correct the data from their effects. Therefore, the models derived from inverse modelling are inevitably biased.

The approach used during my PhD is called forward modelling: the parameters of the evolution history of galaxy populations (i.e. spectral energy distributions, luminosity functions, and their respective evolution) are optimized through Bayesian inference in order to reproduce the diversity of galaxies in the surveys. We start from a randomly initialized model of galaxy evolution (Stuff, Pegase) and we use a generator of images (SkyMaker) to simulate realistic deep fields. The great advantage of this approach is that the simulated and the observed images contain exactly the same selection biases. We then use simple observables to compare the simulations to the observations. Because of the high dimensionality of the problem, we use a deep learning method to reduce the dimension of the images in order to keep only the essential infirmation: the neural network algorithm developped by Charnock et al. (2018) which seeks to maximize the Fisher information of the images. This allows us to understand the statistical impact of each parameter of our model on the information content of the images, and to optimize the model of galaxy evolution. We are using simulated data to show that this approach is unbiased.

vendredi 26 avril 2019 - 16:00
Salle Entresol Daniel Chalonge, Institut d'Astrophysique
Page web du journal-club / Journal-club's webpage