Neural networks are used in this research intending to perform the hydraulic modeling process in an reverse direction. That is, obtaining pipe's roughness from pressures and flow rates. At the beginning, system's training is done with previously calculated input — output pairs. The training is done on neural networks, which are specially designed for each water distribution network. Opened, closed and of different sizes water distribution networks are used. Several neural architectural configurations are tested. Some aggregation levels are tested too.