Air velocities induced by underground vehicles in metro tunnels equipped with ventilation shafts are obtained using artificial neural networks. Complex tunnel shaft-systems with any number of tunnels and shafts and with most of the practically possible geometries encountered in underground structures can be simulated with the proposed method. A single neural network, of type feed-forward back propagation, with a single hidden layer is trained for modeling a single tunnel segment. Train and tunnel parameters that have influence on the vehicle induced flow characteristics are used together to obtain non-dimensional input and target parameters. First input parameter is the major head loss coefficient of tunnel, fL/D. Blockage ratio, A Train/ATunnel and train aspect ratio, (L/D) Train are selected to be nondimensional input parameters to represent the system geometry. As the final input parameter, skin friction coefficient of the train, Csf; drag coefficient of the train, CD; frontal area of the train, ATrain and lateral area of the train, ALateral are combined into a single overall drag coefficient based on the train frontal area. Non-dimensional VAir/VTrain speed ratio is selected to be the only target parameter.