The general framework of the thesis is the development of hybrid systems combining physical modeling and statistical and neuronal modeling. The subject concerns the modeling of complex physical phenomena, which concern the dynamics of ocean circulation, which are components of climate models. The objective is to model dynamic systems, based on statistical models based on deep neural networks which integrate knowledge and constraints from the physics of the phenomenon. The subject requires in-depth skills in statistical learning and neural networks and an interest in climate modeling. This doctoral project is presente by Marie Deschelle under the co-supervision of Pr P. Gallinari Pr (LIP6), Dr M. Levy and Pr S. Thiria (LOCEAN laboratory).