Abstract : The prediction of the behavior of complex structures in which at least two very different scales can appear, in space as well as in time, require the development of enhanced numerical methods. The starting point of this work is a multiscale computational strategy recently proposed at the LMT-Cachan. It can be seen as a mixed domain decomposition method including space and time homogenization without having the limitations of the classical homogenization techniques. The aim of this PhD is to make the method more efficient and more robust in the context of tridimensional complex problems (viscoelasticity with frictional contact). An automatic enrichment technique of the time macro base is proposed in order to reach numerical scalability which can be partially lost in some case of non linearities. A new adaptive time-space approximation technique (based on the radial approximation) is introduced in the framework of internal variables material models, in order to reduce the amount of computations and increase the robustness of the approach. Finally, some numerical tools are proposed in order to minimize the problem sizes and to handle more precisely the enhanced approximation techniques used in this work. The proposed contributions are exemplified on tridimensional problems simulated with a dedicated finite element software implemented during the PhD.