Abstract : In dentin restoration, collagen fiber network infiltration is an issue. In this thesis, an experimental study is lead prior to numerical modelling. Demineralized dentin porosity is a key parameter in dentin bonding that will inﬂuence the hybrid layer quality. Its characterization could be helpful to improve the monomers inﬁltration. So, the objectives of the experimental study were to assess demineralized dentin porosity and quantify the different porous features distribution within the material using mercury intrusion porosimetry (MIP) technique. We compared hexamethyldisilazane (HMDS) drying and lyophilization (LYO) (freeze-drying) in sample preparation. The results showed two types of pores corresponding either to tubules and microbranches or to inter-ﬁbrillar spaces created by demineralization. Global porosity varied from 59% (HMDS-dried samples) to 70% (freeze-dried samples). Lyophilization drying technique seems to lead to less shrinkage than HMDS drying. FESEM revealed that collagen ﬁbers of demineralized lyophilized samples are less melted together than in the HMDS-dried samples. Using our experimental data , we have constructed a relevant numerical geometrical model of the network. The specificity of our model is that the fibers are taken into account implicitly using a regularized Heaviside function. This function is either used to set the viscosity or to localize the contact line where capillary forces are applied. A level set technique with respect to fluid infiltration front tracking in five fiber networks using the level set method and Navier-Stokes equations with capillary terms is used to point out efficient critical infiltration parameters. A variational formulation which can be implemented in FEM is proposed both for the infiltration front and the contact line. Because of lack of knowledge on fiber orientation, different configurations were tested through permeability assessment of the whole network. Fiber orientation, interfibrillar space and contact angle influence were investigated. Then a network representing the dentinal porous substrate is infiltrated by a wetting fluid and we came to a useful conclusion : increasing infiltration time will not improve the quality of the bonding. In order to link the numerical and experimental studies, infiltration curves of mercury into the porous network are compared. We used a non-wetting fluid (mercury) to invade the numerical porous network to model the porosimetry test. As mercury do not invade the porous network naturally, an incremental pressure is applied to force it to penetrate the porous medium. The simulation is qualitatively satisfying and enables to point out the porosimetry test limitations in a pedagogical way. Thanks to the simulation, we can see the mercury invading pore spaces: big pores are penetrated first and then small pores (which obviously cannot be seen during the experimental process). Quantitatively, two sizes of pore are detected and their contributions to total porosity are satisfying. Nevertheless, the pore size values have to be ajusted. To conclude, this thesis contributed to the characterization of demineralized dentin. The numerical modeling allows us to point out influencing parameters during capillary infiltration of a fibrous network and to illustrate the porosimetry test process an its limitations. An interesting tool to confirm a geometrical model has been built. The tools that have been built can easily been used to other porous structures.