Achieving an optimal design for a typical water distribution system (WDS) essentially involves determining which combination of pipes and arrangements will produce the most efficient and economical network. Solving the problem is a complex process, one well suited to computationally intensive heuristic methods. Including water quality constraints can pose a special challenge due to the demanding, extended-period simulations involved. Employing artificial neural networks (ANNs) can reduce the amount of computation time needed. ANNs can in fact approximate disinfectant concentrations in a fraction of the time required by a conventional water quality model. This study presents a methodology for improving the accuracy of ANNs applied to the optimal design of a WDS by means of a probabilistic approach based on the fast finding of a network similar to the optimal WDS. This work also presents a methodology to find such a network. ANNs trained with the probabilistic dataset generated using the proposed approach were shown to be more accurate than their counterparts trained with a random dataset.