One of the major scientific problems in seismology is to estimate the ground motions expected at a given site from a future earthquake. The aim of this thesis is to test and validate two different methods of ground motions simulation based on the empirical Green's function (EGFs) approach and to provide elements that can help to develop a blind simulation methodology. In a first part, a simulation method based on a stochastic point source approach is validated on the real data of recent earthquakes well ins- trumented. Taking advantage of the availability of good quality records of small earthquakes that can be used as EGFs, two applications are made : one application to the Les Saintes earthquake Mw = 6.4 and one application to the L'Aquila earthquake Mw = 6.3. We have developed a blind simulation approach by taking into account an uncertainty on the parameter of stress drop ratio C. This approach allows to generate a set of synthetic accelerograms of a target earthquake varied enough to be representative of a large number of possible source scenario and taking into account in a statistical sense potential direc- tivity effects. Moreover, the variability of ground motions produced by our blind simulation approach is consistent with the variability of ground motion predictions purely empirical. Putting themselves in a real context of blind simulation, this approach is also applied to the simulation of an historical Pyrenean earthquake Mw = 6.1. In a second part, to be closer to current knowledge on the complexity of the kine- matics of a seismic source, we use a more complex extended source model, combining kinematic models of fractal composite sources with EGF approach to produce broadband synthetic accelerograms. The po- tential of the method is tested on an application to L'Aquila earthquake by setting the input parameters providing the best opportunity to reproduce the recordings of the target earthquake. This has produced very satisfying results on all ground motion parameters analyzed. This simulation method appears to be very promising for the implementation of a blind simulation methodology, even if the main difficulty lies in the need to define the variability of many poorly known input parameters in the simulation of a future earthquake.
