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

Area of Science:

• Environmental Engineering
• Hydrogeology
• Computational Science

Background:

• Pump-and-treat (P&T) is a key groundwater remediation technology.
• Optimizing P&T operations (well locations, pumping rates) is crucial for efficiency and cost reduction.
• Surrogate models coupled with optimization algorithms are emerging tools for P&T scheme formulation.

Purpose of the Study:

• To compare the performance of five surrogate models (Kriging, Polynomial Interpolation, SVR, RF, DNN) for predicting contaminant removal in P&T remediation.
• To explore the potential of combining multiple surrogate models for enhanced P&T optimization.
• To develop and evaluate a multi-surrogate optimization framework for groundwater remediation.

Main Methods:

• Evaluated five surrogate models: Kriging, Polynomial Interpolation, Support Vector Regression (SVR), Random Forest (RF), and Deep Neural Network (DNN).
• Assessed model predictive accuracy for contaminant removal efficiency across diverse P&T schemes.
• Developed a multi-surrogate optimization framework integrating all five models with a genetic algorithm.

Main Results:

• Deep Neural Network (DNN) showed the highest overall prediction accuracy, but no single model was consistently superior across all cases.
• A multi-surrogate framework combining diverse models leveraged complementary strengths.
• The optimized remediation schemes achieved superior contaminant removal (17.5% residual) compared to non-optimized schemes (19.2-21.7%).

Conclusions:

• Combining multiple surrogate models offers significant benefits for optimizing pump-and-treat remediation schemes.
• The developed multi-surrogate framework provides a robust tool for environmental management.
• This study advances surrogate-based optimization techniques for groundwater cleanup.