Approaching the inverse problem of parameter estimation in groundwater models by means of artificial neural networks

Abstract

Mathematical models are widely used to simulate the behaviour of groundwater systems under various physical conditions. The development of these models requires that the system state evolution be predicted by solving appropriate governing equations. This also requires that the values of the parameters imbedded in the models be accurately determined. Since the hydraulic and water quality parameters are in general not directly measurable, their estimation must be obtained by solving the so called inverse parameter identification problem which amounts to determining the unknown physical parameters by fitting observed histories of the system states. In this paper we investigate the feasibility of solving the inverse problem by using artificial neural networks. These computational tools have the ability to learn complex input-output relationships and have a very advantageous property of generalization. To demonstrate the approach, a simple case study is considered consisting of an analytic model of contaminant transport due to a point source in stationary flow field. The model is used to simulate the behaviour of the groundwater systems for different values of the dispersion coefficient. A set of supervised, multilayered, feedforward neural networks are then trained to predict the values of the parameter corresponding to given concentration histories.