Optimized stacking, a new method for constructing ensemble surrogate models applied to DNAPL-contaminated aquifer...

Area of Science:

• Environmental Engineering
• Computational Hydrogeology
• Machine Learning Applications

Background:

• Dense non-aqueous phase liquid (DNAPL) contamination poses significant challenges for aquifer remediation.
• Surfactant-enhanced aquifer remediation (SEAR) is effective but costly, necessitating optimization.
• Numerical simulations for SEAR optimization are computationally intensive, requiring efficient alternatives.

Purpose of the Study:

• To develop and evaluate an accurate surrogate model for optimizing SEAR strategies.
• To compare the performance of the Stacking ensemble method against conventional machine learning and ensemble techniques.
• To identify the most cost-effective DNAPL remediation strategy using the developed surrogate model.

Main Methods:

• Employed six machine learning methods as base surrogate models with various feature scaling techniques.
• Utilized Bagging and Boosting homogeneous ensemble methods to enhance base model accuracy.
• Implemented a Stacking ensemble method, incorporating 18 base models, and optimized using Bayesian hyper-parameter optimization.
• Integrated the optimized Stacking surrogate model into a differential evolution optimization algorithm.

Main Results:

• The artificial neural network model with power transformer scaling achieved the best standalone performance (CV RMSE: 0.065).
• Boosting homogeneous ensembles improved base model accuracy, with the best model achieving a test RMSE of 0.039.
• The Stacking ensemble model significantly outperformed other methods, reaching a cross-validation RMSE of 0.016.
• Bayesian hyper-parameter optimization surpassed random and grid search for model tuning.

Conclusions:

• The Stacking ensemble surrogate model provides a highly accurate and computationally efficient alternative to numerical simulations for SEAR optimization.
• The developed approach successfully identified an optimal remediation strategy with a total cost of $72,706.
• This study demonstrates the potential of advanced machine learning techniques for cost-effective environmental remediation.