Abstract. We present a methodology to account for the stochastic nature of hydraulic conductivity during the design of pump-and-treat systems for aquifer cleanup. The methodology (1) uses a genetic algorithm to find the global optimal solution and (2) incorporates a neural network to model the response surface within the genetic algorithm.

Abstract. We present and apply a new simulation/optimization approach for single-and multiple-planning period problems in groundwater remediation. Instead of the traditional control locations for contaminant concentrations, \Ve use an LC>O norm as a global measure of aquifer contamination (CMAX). We use response-surface constraints to represent CMAX within the optimization model. We compare the performance of formal mixed integer nonlinear programming and a genetic algorithm for several optimization scenarios.

No abstract

The Department of Defense (DoD) Environmental Security Technology Certification Program and the Environmental Protection Agency sponsored a project to evaluate the benefits and utility of contaminant transport simulation-optimization algorithms against traditional (trial and error) modeling approaches. Three pump-and-treat facilities operated by the DoD were selected for inclusion in the project. Three optimization formulations were developed for each facility and solved independently by three modeling teams (two using simulation-optimization algorithms and one applying trial-and-error methods). The results clearly indicate that simulation-optimization methods are able to search a wider range of well locations and flow rates and identify better solutions than current trial-and-error approaches. The solutions found were 5% to 50% better than those obtained using trial-and-error (measured using optimal objective function values), with an average improvement of~20%. This translated into potential savings ranging from $600,000 to $10,000,000 for the three sites. In nearly all cases, the cost savings easily outweighed the costs of the optimization. To reduce computational requirements, in some cases the simulation-optimization groups applied multiple mathematical algorithms, solved a series of modified subproblems, and/or fit ''meta-models'' such as neural networks or regression models to replace time-consuming simulation models in the optimization algorithm. The optimal solutions did not account for the uncertainties inherent in the modeling process. This project illustrates that transport simulation-optimization techniques are practical for real problems. However, applying the techniques in an efficient manner requires expertise and should involve iterative modification to the formulations based on interim results.

No abstract