The aim of this thesis is to identify neural network structures capable of reducing the algorithmic cost of solving a CFD system over long periods, in complex domains in dimension 2 or 3, while guaranteeing quantitative control of the resolution error. More precisely, the aim is to design a network acting as a numerical scheme, enabling us to evaluate an approximation of the system variables at instants t, based on the values of these variables at an initial instant t0. As the database will be relatively limited to a few numerical resolutions of these equations, and insofar as we wish to guarantee certain properties on the approximation obtained, it is essential to adapt the network structure to explicitly integrate the physics of the problem. To this end, the aim is to develop an approach that combines physical elements with a classical network structure.