A multiple surrogate simulation-optimization framework for designing pump-and-treat systems

Abstract

Pump-and-treat (P&T) remediation is a widely adopted and effective method for groundwater contamination control. It is important to optimize the operation schemes (pumping well locations and pumping rates) to maximize contaminant removal efficiency and minimize operational costs. Recently, surrogate models have been integrated with optimization algorithms to formulate the remediation schemes. However, with various surrogate techniques available, their comparative performance in P&T remediation tasks and potential for combined usage of multiple surrogates require further exploration. In this study, five popular surrogate models-Kriging, Polynomial Interpolation, Support Vector Regression (SVR), Random Forest (RF), and Deep Neural Network (DNN)-were evaluated for their ability to predict contaminant removal efficiency under diverse schemes in a multi-contaminant site. The analysis revealed that, while DNN achieved the highest overall prediction accuracy in the validation stage across the 200 cases, no single surrogate model consistently outperformed the others in all individual cases. A multi-surrogate optimization framework, coupling all five models with a genetic algorithm, was developed to enhance P&T schemes. The usage of multiple surrogates finally brings benefits because the complementary strengths of diverse surrogate models are combined. We identified remediation schemes that achieved superior contaminant removal (17.5% residual contaminant) compared to the other results (19.2-21.7%). The framework offers a robust tool for environmental management and insights for advancing studies related to surrogate-based optimization.

Keywords: Genetic algorithm; Groundwater contaminant remediation; Machine learning; Multi-model framework; Pump-and-treat optimization; Surrogate modeling.