Concrete durability is strongly affected by the flow of fluids, gas and pollutants in its porous matrix. The presence of cracks weakens the resistance of concrete porous matrix and constitutes preferential flow paths for aggressive components. In the thesis, a probabilistic numerical modeling strategy for modeling fluids transfers in cracked concrete structures is presented. The concrete is modeled in the framework of water saturated porous media. Its (mechanical) cracking is modeled by means of a macroscopic probabilistic approach, explicitly taking into account material heterogeneity as well as size effects. The main assumption of the model, developed in the frame of the the Finite Element Method, is to consider a finite element volume as a volume of heterogeneous material and to assume that physical mechanisms influencing the cracking processes remain the same whatever the scale of observation. At the scale of the finite element, mechanical properties are then functions of its own volume. To describe the heterogeneity of the material, these mechanical properties are consider as uncorrelated random variables distributed over the finite element mesh. Characteristics of statistical distribution laws are directly depending on the degree of heterogeneity of the finite element (the ratio between its volume and the volume of the coarsest aggregate) and of the quality of the cement paste. An inverse analysis approach allows to find their parameters as functions of the elementary volume. A weak coupling between cracking and fluid transfers is considered, under the assumption of no interaction between the two processes (i.e. the mechanically produced cracking of a finite element induce a local variation of its permeability tensor). An experimentally adapted Pouiseuille law, based on an original experimental protocol, allows to relate this permeability variation to the crack aperture and to macroscopically take into account the influence of crack roughness, aperture variation and tortuosity. A Monte-Carlo like approach is used in order to statistically validate mechanical and hydraulic simulations. The coupling strategy is validated in two phases, both at the scale of a laboratory specimen and at the scale of a multi-cracked structural element
