Developing generic hydrodynamic models using artificial neural networks

Possibilities for the development of a new modelling paradigm, namely allowing models to 'construct themselves' by learning from existing numericalhydraulic models, was investigated by extending previous works to encompass schemes that can be applied over arbitrary bathymetries with variable distances and time steps. For the simplest possible cases of one and two dimensional flow problems considered in this study, the relatively elementary technology of artificial neural network was found to provide acceptable results. Moreover, It was demonstrated that the well-trained networks could be substituted in place of the finite difference schemes in the hydrodynamic model formulation and could perform like numerical operators. This new paradigm is intended in future to supplement, and even in some instances to replace the current one