Comparison of Surrogate Models Based on Different Sampling Methods for Groundwater Remediation

Abstract

To assess the influence of sampling methods on surrogate models’ accuracy, using two test problems and a nitrobenzene-contaminated aquifer remediation problem, several sampling methods were adopted to collect sample data sets and a Kriging method was adopted to construct surrogate models. The sampling methods adopted include Latin hypercube sampling (LHS), space-filling-based LHS (SFLHS), orthogonal-array-based LHS (OALHS), and space-filling and orthogonal-array-based LHS (SFOALHS). The space-filling properties and orthogonality of sampling results, as well as the corresponding surrogate models’ accuracies, were compared, and the best surrogate model was invoked for assessing the remediation efficiency in a groundwater remediation optimization problem. The results indicated that (1) compared with LHS, SFLHS, and OALHS results, the SFOALHS result had the best trade-off between orthogonality and space-filling property, and better represented the population; and (2) the SFOALHS-based surrogate model was more accurate and better fit the simulation model in both test problems and the case study; therefore it was invoked as a constraint condition for replacing the behavior of the computational simulation model in the optimization process.