Optimizing groundwater remediation designs using dynamic meta-models and genetic algorithms

Large-scale water resources optimization often involves using time-intensive simulation models to evaluate potential water resource designs or calibrate parameter values. The extensive computational requirements have become a critical issue for applying optimization algorithms to water resources problems. In order to improve the computational efficiency, this research developed a dynamic modeling framework called Adaptive Meta-model Genetic Algorithm (AMGA), in which time-efficient meta-models are adaptively retrained within a genetic algorithm (GA) to replace time-intensive simulation models. Different configurations of AMGA were tested on two groundwater remediation design cases. In these applications, AMGA saved 85-90% percent of the simulation model calls with little loss in accuracy of the optimal solutions.; The prediction accuracy of meta-models is a key factor in determining AMGA's performance. This research then developed an advanced meta-model construction method to achieve improved predictions. In this research a trust region-based meta-model approach, in which a global model and a set of local models are constructed into a hierarchical ensemble, is developed. The technique was tested on a remediation case study. The results show that the adaptive GA coupled with the trust region-based meta-models converged with somewhat higher accuracy and identified the optimal solutions faster. The technique was also tested on a nitrate concentration prediction problem. The results show that the trust region-based prediction models had better performance than the corresponding single-prediction models.; Real-world optimization problems are often inherently uncertain. The last focus of the research is to extend the adaptive modeling technique in a stochastic optimization framework so that robust optimal solutions can be efficiently identified in the presence of parameter uncertainty. The developed algorithm, called Noisy-AMGA, minimizes the expected fitness function with a constrained reliability level. As in AMGA, the meta-models in Noisy-AMGA are online updated but they are trained to predict the expected outputs. The method was applied to two remediation case studies, where the primary source of uncertainty stems from hydraulic conductivity values in the aquifers. The results show that the technique can lead to far more reliable solutions with significantly less computational effort.