Solving the inverse problem of parameter estimation by genetic algorithms: the case of a groundwater contaminant transport model

Mathematical models are widely used within the performance assessment of radioactive waste repositories to describe the behaviour of groundwater systems under the various physical conditions encountered throughout the long time scales involved. The effectiveness of such predictive models largely depends on the accuracy with which the involved parameters can be determined. In the present paper, we investigate the feasibility of using genetic algorithms for estimating the parameters of a groundwater contaminant transport model. The genetic algorithms are numerical search tools aiming at finding the global optimum of a given real objective function of one or more real variables, possibly subject to various linear or non linear constraints. The search procedures provided by the genetic algorithms resemble certain principles of natural evolution. In the case study here, the transport of contaminants through a three-layered monodimensional saturated medium is numerically simulated by a monodimensional advection-dispersion model. The associated velocity and dispersivity parameters are estimated by a genetic algorithm whose objective function is the sum of the squared residuals between pseudo-experimental data, obtained with the true values of the parameters, and the concentration profiles computed with the model using the estimated values of the parameters. The results indicate that the method is capable of estimating the parameters values with accuracy, also when in presence of substantial noise. Furthermore, we investigate the possibility of extracting some qualitative information regarding the sensitivity of the model to the unknown input parameters from the speed of convergence and stabilization of the identification procedure. (C) 2002 Elsevier Science Ltd. All rights reserved