Les simulations, de plus en plus, sont capables de saisir la complexité de l'époque de réionisation, durant laquelle l'hydrogène neutre de l'Univers a été ionisé par les premières sources lumineuses. Des bases de données représentatives de la multitude de signaux possibles seront nécessaires pour contraindre les paramètres des modèles quand des observations 21~cm seront disponibles. À cette fin, et en préparation des observations à venir sur des instruments comme SKA, nous avons développé une base de données de cones de lumières EoR haute-résolution (21ssd.obspm.fr), ainsi qu'une modélisation du bruit thermique. Nous avons également développé un formalisme permettant de quantifier la différence entre les modèles de cette base de données, en utilisant le spectre de puissance et la fonction de distribution des pixels. Nous trouvons que les deux diagnostiques sont sensibles à des paramètres différents des modèles, ce qui signifie que les deux peuvent être utilisés ensemble de manière complémentaire pour extraire l'information maximale. De plus, en utilisant le code 21cmFAST, nous avons développé des stratégies pour échantillonner l'espace des paramètres d'une manière optimale (plus homogène et isotrope), afin de fournir le meilleur point de départ entrainer un réseau de neurones. Ce réseau retrouve les paramètres du modèle en se basant sur une observable. Nous observons une amélioration modérée dans la précision de ses prédictions quand nous utilisons l'échantillonnage optimisé lors de son entrainement. / Simulations are increasingly able to capture the intricacies of the Epoch of Reionization, during which the neutral hydrogen in the Universe was ionized by the first luminous sources. Databases encompassing the range of possible signals will be needed to constrain parameter values when 21~cm observations are available. In preparation for upcoming experiments such as the SKA, we have developed a database of high-resolution EoR lightcones (21ssd.obspm.fr), along with realistic thermal noise modelling. We examine frameworks with which we can quantify the difference between entries in this database, specifically with the power spectrum and pixel distribution function. We find that the two diagnostics are sensitive to different parameters, meaning they can be used together to extract maximumal information. Then, using the 21cmFAST code, we explore how to optimally sample a parameter space (so that it is more homogeneous and isotropic), in order to provide the best set-up for parameter extraction. Finally, the improved sampling is used in training a neural network. The neural network uses observables as input data, and attempts to estimate the corresponding parameter values. When the optimal sampling is used as training data, we find that the neural network is able to estimate parameter values with a modest improvement in accuracy.

In the standard model of cosmology (ΛCDM) the first stars, star clusters, and galaxies are expected to have formed in short bursts of star formation in low-mass dark matter halos at high redshifts (<img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?z%5C,%20%5Csim%20%5C,6-10" data-classname="equation_inline" />). Up to this point, attempts to predict the properties and abundances of these luminous objects have made use of numerically expensive cosmological simulations. On top of being numerically expensive, these simulations often lack the required sub-parsec resolution needed to resolve the formation of compact star clusters and/or neglect possibly dominant stellar feedback processes. Motivated by this, I introduce Anaxagoras, as far as I know the most detailed analytical ab initio model of starbursts in low-mass halos to date. The model incorporates sub-models for gas cooling (including a new determination of the H2-cooling threshold in minihalos), central gas accretion and disk formation (using a new selfsimilarsolution), stellar feedback from radiation pressure (direct stellar radiation, Lyman-<img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?%5Calpha" data-classname="equation_inline" data-title="" /> scattering in H I, and multiple scattering of IR photons by dust), stellar winds, expanding HII regions, and (crudely) supernovae. The resulting star formation efficiency is used to predict the fraction of stars that remain gravitationally bound in a cluster following gas expulsion, andwhat fraction escape the central region of the halo, yet remain bound by the dark matter halo. I apply Anaxagoras to study star formation at <img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?z%5C,%20%3E%20%5C,6" data-classname="equation_inline" /> in satellite halos of the Milky Way using a halo merger tree code, as well as Population III (Pop III) star formation in minihalos. For the Milky Way setup, hundreds of galaxies are predicted to form with luminosities (<img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?L_%7B%5Crm%20V%7D%20%5C,%3C%5C,%20%5Crm%7Bfew%7D%5C,%20%5Ctimes%20%5C,%2010%5E4%20%5C:%20%5Crm%7BL%7D_%7B%5Codot%7D" data-classname="equation_inline" />), half-mass radii (<img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?%5Csim%2010-200%5C:%5Crm%20pc" data-classname="equation_inline" />), mass-to-light ratios (<img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?M/L_%7B%5Crm%20V%7D%20%5Csim%20100%20-%20%5Crm%7Bfew%7D%20%5C,%5Ctimes%5C,%2010%5E3%20%5C:%5Crm%7BM%7D_%7B%5Codot%7D/%5Crm%7BL%7D_%7B%5Codot%7D" data-classname="equation" />), and ages (<img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?13.18%5E%7B+0.29%7D_%7B-0.31%7D%5C:%5Crm%20Gyrs" data-classname="equation_inline" data-title="" />) in good agreement with the observed local population of Ultra-Faint Dwarfs. This shows that ΛCDM is able to explain the properties ofthe faintest dwarf galaxies without fine-tuning. Furthermore, at least ~ 40 compact (initial half-mass radii <img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?%5Csim%200.1-5%5C;%5Crm%7Bpc%7D" data-classname="equation_inline" />), old (<img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?13.27%5E%7B+0.21%7D_%7B-0.39%7D%5C:%5Crm%20Gyrs" data-classname="equation_inline" />) globular cluster (GC) candidates with initial stellar masses <img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?10%5E5%20-%2010%5E6%20%5C:%20%5Crm%7BM%7D_%7B%5Codot%7D" data-classname="equation_inline" />  are predicted to form at the center of low-mass halos, and could survive to the present-day and explain at least a fraction of the observed metal-poor GCs. Their properties are consistent with recent candidates for GCs residing in dark matter halos. Thus, Anaxagoras lends support to the viability of the scenario of GC formation in minihalos. Finally, the formation of Population III (Pop III) stars in minihalos is studied, with the conclusion that if Pop III stars are not overly massive (<img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?25%5C:%5Crm%7BM%7D_%7B%5Codot%7D" data-classname="equation_inline" />) between ~ 1 − 30 stars could form per minihalo at <img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?z%5C,%20%3E%20%5C,20" data-classname="equation_inline" />, with the number increasing to ~ 10 − 500 stars per minihalo at <img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?z%5C,%3C%5C,15" data-classname="equation_inline" /> as Lyman-Werner feedback delay star formation until halos reach larger masses. In the case where Pop III stars are more massive (<img src="http://www.diva-portal.org/cgi-bin/mimetex.cgi?140%5C:%5Crm%7BM%7D_%7B%5Codot%7D" data-classname="equation_inline" data-title="" />) most minihalos form just a single star. Due to self-shielding of H2 in minihalos, I find that the cosmological Lyman-Werner background is insufficient to produce Pop III galaxies in atomic-cooling halos, with the implication that the number of massive Pop III galaxies/star clusters in the early Universe has been greatly overestimated in the literature that ignores self-shielding.