We have implemented a neural network of three hidden layers with 40 neurons each layer to be used as soil/rock transfer functions for two stations in Mexico City. The net was trained with supervised learning through input and output vectors of accelerations (twelve records, from five seismic events from Guerrero and Puebla, 5.8 M 7.3), and tested with three records not taken in account in the training. The results in the frequency domain are good, finding a seismic amplification between 0.2 to 5 Hz for the Lake zone. In the time domain we obtain results that are not coincident. Due to the data and the complex of the phenomena, it is necessary to apply this tool using more records for the training net, so the phenomena can be learned better through reliable database.