Forecasting stream-aquifer interactions using innovativemethods and integration into contaminant transport models

Groundwater constitutes a substantial portion of global freshwater reserves and plays a critical role in the hydrological cycle due to its relatively low vulnerability to contamination. However, in regions where it interacts with rivers, surface water–groundwater interactions exhibit considerable complexity, governed by factors such as topography, regional flow regimes, and hydraulic head gradients. This multifactorial nature has increased interest in exploring innovative methodologies for the quantitative assessment of such interactions. In this context, conducting preliminary evaluations on controlled, theoretically derived scenarios is considered a valuable step toward assessing the predictive capabilities and physical consistency of emerging modeling tools. This thesis investigates the applicability of machine learning–based and empirical approaches for estimating interaction flow rates and characterizing selected parameters related to contaminant transport in river–aquifer systems. In the first phase, data generated from scenarios involving a sloped river interacting with an adjacent aquifer were analyzed using XGBoost, CatBoost, Random Forest, and LightGBM algorithms. The interaction flow rate was defined as the target variable, and model performances were compared, including an evaluation of how different data generation processes affect predictive accuracy. Additionally, a new empirical equation was developed based on relevant physical parameters. In the second phase, the travel time of a particle from the aquifer to the river was predicted using the XGBoost algorithm, where feature engineering notably improved model performance. In the third phase, the advection–dispersion equation was employed to analyze mass flux and total contaminant load under varying scenarios. Using the resulting datasets, the time at which a contaminant transported through the aquifer would generate a maximum concentration in the river was predicted via the CatBoost algorithm. The results indicate that both machine learning–based and empirical methods are promising tools for modeling surface water–groundwater interactions and highlight the importance of data preprocessing and feature selection in enhancing model performance.