Semi-analytical models (SAMs) are currently the best way to understand the formation of galaxies and clusters within the cosmic web dark-matter structures. While they fairly well reproduce the local stellar mass function, correlation function and luminosity function, they fail to match observations at high redshift (z>3) in most cases, particularly in the low-mass range. The inconsistency between models and observations indicates that the history of gas accretion in galaxies, within their host dark-matter halo, and the transformation of gas into stars, is not well followed. In this thesis, we introduce a new version of the GalICS model and we use it to explore the impact, on the star formation and in the low-mass range, of supernovae feedback and photo-ionization. These two mechanisms are commonly used to limit the amount of gas available to form stars.We will show that, even with a strong efficiency, these two process cannot explain the observed stellar mass function, luminosity functions, and the stellar mass versus dark matter halo mass relation. We will thus introduce two ad-hoc modifications of the standard paradigm. We propose first a strong modification of the star formation efficiency as a function of the dark matter halo mass. This model produces good results, especially on the faint end of the stellar mass function, but is, by construction, in disagreement with the well known Kennicutt star formation law. We will thus introduce a deeper change, based on a no star-forming gas component, and a new gas distribution in the galaxy discs. The reservoir in which stays the no star-forming gas generates a delay between the gas accretion and star formation. This model is in very good agreement with a large set of observations. However, it poses the question of the origin of the no star-forming gas. We will discuss its origin in the framework of the large scale disturbed dynamic of high-redshift structures.
