During the last years, neural networks applications have been disseminated in many knowledge areas, including civil engineering. In the middle 90`s, a research work had been started in Brazil, in order to investigate the efficiency of neural networks in the analysis of soil behavior and problems involving geotechnical engineering. This thesis is the result of part of these studies, where some potentialities of neural networks technique are presented. Three different feedforward NNs applications in geotechnical engineering are presented. Levenberg- Marquardt algorithm was used for training. The first application is the simulation of results of dynamic pile tests, obtained from CAPWAP analysis, showing that it is possible to do a field pre-analysis of the pile behavior, which is still unpracticable when the traditional CAPWAP method is used. The second application is related to the study of two different soils behavior:sand from Ipanema and residual gnaissic soil from Rio de Janeiro. Results of submerged and non submerged direct shear tests and drained and undrained triaxial compression tests were used. The third application involves the simulation of subsoil characteristics of Angra 2 Nuclear Power Plant site. The available information came from SPT bulletins. Simulations involving several types of soil layers spatial distribution, water level position, penetration strength of soils and local topography were performed. The obtained results were very satisfactory. It can be concluded that the neural networks technique presents great applicability in resolution of geotechnical problems with different characteristics, showing an efficiency as good or even better than other traditional numerical techniques.