An adaptive dynamic surrogate model using a constrained trust region algorithm: application to DNAPL-contaminated-groundwater-remediation design

Abstract

Simulation optimization is a robust tool in the design of groundwater remediation strategy. Simulation models for groundwater contaminated by dense non-aqueous phase liquids (DNAPLs) are usually computationally expensive, especially when used within a simulation optimization framework. Surrogate models have been proved to possess the potential to speed up complex models without sacrificing accuracy or detail. In this study, a constrained trust region (CTR)-based adaptive dynamic surrogate model was proposed and applied to a DNAPL-contaminated-groundwater-remediation-design scenario. First, the multiphase flow simulation model was constructed to simulate the groundwater remediation process. Then, preliminary input–output sample points were generated using the Latin hypercube sampling method and the developed simulation model. Next, the static Kriging surrogate model, as well as the nonlinear optimization model, were constructed. Finally, the surrogate model and optimization model were updated based on the CTR method and the optimal remediation strategy was obtained. The application of the CTR-based dynamic surrogate model reduced the mean relative error by 46% compared with the static surrogate model, and greatly improved the reliability of the optimal remediation strategy. The CTR-based dynamic surrogate model had the best fitting accuracy of multiphase flow simulation when compared with the expected improvement-based dynamic surrogate model and the optimal solution-based dynamic surrogate model. The corresponding optimization model also had the minimum remediation cost. This study demonstrates that the proposed CTR-based dynamic surrogate model is an effective tool to increase the accuracy of surrogate models and the reliability of their optimization results.