A Genetic Programming-Based Surrogate Model Development and Its Application to a Groundwater Source Identification Problem

This paper investigates a groundwater source identification problem in which chemical signals at observation wells are used to reconstruct the pollution loading scenario. This inverse problem is solved using a simulation-optimization approach that uses a genetic algorithm to conduct the search. As the numerical pollution-transport model is solved iteratively during the heuristic search, the evolutionary search can be in general computationally intensive. This is addressed by constructing a surrogate modeling approach that is able to predict quickly the concentration profiles at the observation wells. A genetic program is used in the development of the surrogate models that provides an acceptable prediction performance. The surrogate model, which replaces the numerical simulation model, is then coupled with the evolutionary search procedure to solve the inverse problem. The results will illustrate 1) the performance of the surrogate model in predicting the concentration compared with the predictions using the original numerical model, and 2) the quality of the solution to the inverse problem obtained using the surrogate model to that obtained using the numerical model.