Groundwater remediation design using multiscale genetic algorithms

Water resources optimization models often use spatial numerical models to approximate the physics of natural systems. The discretization of the numerical grids can affect their search for optimal solutions, in terms of both solution reliability and computational costs. Computational costs are particularly significant for population-based optimization techniques such as genetic algorithms (GAs), which are being applied to water resources optimization. To overcome these bottlenecks, this paper proposes multiscale strategies for GAs that evaluate designs on different spatial grids at different stages of the algorithm. The strategies are initially tested on a hypothetical groundwater remediation problem, and then the best approach is used to solve a field-scale groundwater application at the Umatilla Chemical Depot in Oregon. For the Umatilla case, the multiscale GA was able to save as much as 80% of the computational costs (relative to the GA that used only the fine grid) with no loss of accuracy, thus exhibiting significant promise for improving performance of GA-based optimization methodologies for water resources applications.