In this work it is introduced a methodology for the supervised training of the Neural Networks, using Levenberg Marquardt's algorithm. This method is applied in gravimetry for the optimization in the convergence at the inverse problem.