Saltwater scavenging optimization under surrogate uncertainty for a multi-aquifer system

Abstract

Surrogate-assisted simulation-optimization technique is an effective approach for identifying aquifer remediation strategies. Using surrogate models to replace high-fidelity groundwater simulation models is key to reduce computing time. Inevitably, uncertainties persist in surrogate model structure and parameters and may lead to poor accuracy in prediction. However, past studies often adopt one surrogate to capture the original model response and rarely investigate uncertainties and approximation accuracy of surrogate models. In this study, we present an ensemble surrogate modeling approach to address uncertainty in surrogate model predictions for saltwater intrusion mitigation designs using a hypothetical horizontal scavenger well. Firstly, polynomial response surface surrogate models are constructed to capture the relationship between chloride concentrations and saltwater extraction rates of a horizontal well. Secondly, a stochastic optimization model that includes multiple ensemble surrogates is developed to determine the optimal saltwater extraction schedule. Chance-constrained (CC) programming is used to account for model selection uncertainty in probabilistic nonlinear concentration constraints. Akaike information criterion correction (AICc) based Bayesian model averaging (BMA) method is adopted to determine required statistics of concentrations in the chance constraints. Finally, a set of saltwater intrusion mitigation strategies are obtained through an interior point optimization algorithm. The proposed methodology is applied to the saltwater intrusion problem in Baton Rouge, Louisiana. A high-fidelity flow and solute transport model is developed, but requires 16 h for one model run. The study shows that the developed surrogate models perform well and have strong predictive capabilities with respect to the solute transport model response, while considerably reducing computation time. Including model selection uncertainty through multimodel inference and model averaging provides more reliable remediation strategies compared with the single-surrogate assisted approach.