Gaussian Process Regression Tuned by Bayesian Optimization for Seawater Intrusion Prediction.

Area of Science:

• Hydrogeology and Computational Modeling
• Machine Learning Applications in Environmental Science

Background:

• Accurate prediction of seawater intrusion is crucial for managing coastal groundwater resources and preventing water quality degradation.
• Traditional variable density models are computationally intensive, limiting their application in scenarios requiring numerous simulations.
• Developing efficient surrogate models is essential to balance prediction accuracy with computational cost.

Purpose of the Study:

• To investigate Gaussian process regression (GPR) as a surrogate model for predicting seawater intrusion extent.
• To compare the performance of GPR against other regression methods, including regression trees and Support Vector Machine regression.
• To assess the efficiency of Bayesian optimization in enhancing the predictive capabilities of these models.

Main Methods:

• Utilized Gaussian process regression, a non-parametric kernel-based probabilistic model, as a surrogate for complex variable density simulations.
• Defined seawater intrusion extent by the 0.5 kg/m³ iso-chlore (seawater intrusion toe) at the aquifer's bottom.
• Employed 4000 variable density simulations with varying pumping rates and initial concentrations, sampled using Latin hypercube sampling, to train the surrogate models. Compared GPR with regression trees and Support Vector Machine regression, optimizing all with Bayesian optimization.

Main Results:

• Gaussian process regression demonstrated superior performance in predicting the seawater intrusion toe compared to regression trees and Support Vector Machine regression.
• GPR achieved lower Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), and a higher coefficient of determination (R²).
• Bayesian optimization effectively enhanced the predictive efficiency across all tested regression methods.

Conclusions:

• Gaussian process regression is a highly efficient and accurate surrogate modeling technique for predicting seawater intrusion extent, despite higher computational demands during training.
• The findings support the use of GPR for improved groundwater management and protection strategies in coastal areas.
• This study highlights the potential of machine learning approaches to overcome the limitations of traditional numerical models in hydrogeological applications.