Predicting Residence Times in Stream-Aquifer Systems with XGBoost Machine Learning Algorithm

Groundwater residence times are key to unraveling the complex dynamics of aquifers, providing insights into their hydrological processes and contaminant transport mechanisms. Hyporheic flow, as an integral component of surface water-groundwater interactions, causes the exchange of substances between surface water and groundwater, enabling pollutants to migrate into aquifers, travel within them, and eventually return to surface water. Groundwater residence time in such systems plays a vital role in developing strategies to protect water resources and promote their sustainable use. In the literature, various analytical and numerical models have been applied to estimate residence time. In addition to these approaches, advancements in technology have introduced artificial intelligence and machine learning methods as valuable tools for determining residence time. The performance of different algorithms in calculating residence time may vary depending on the complexity and specific characteristics of the model. Therefore, investigating the performance of machine learning methods in this context is essential. This study aims to predict the travel times of particles to a stream within a stream-aquifer system using the XGBoost machine learning algorithm. The datasets used in the study were prepared based on a previously developed mathematical model for the system. The velocity vectors derived from the mathematical model were employed to calculate the travel times of particles to the stream. To train the machine learning model and estimate residence time, six parameters affecting travel time were analyzed: hydraulic conductivity (K), stream slope (S), aquifer length (L_y), aquifer width (L_x), and the x and y coordinates of the particles. During model development, random scenarios were generated to create training data. Feature engineering was applied to improve model accuracy, incorporating derived parameters such as “L_y×S” and replacing the x and y coordinates with more meaningful features like the “y/x” ratio. The results demonstrated that hydraulic conductivity and the “L_y×S” parameter had the most significant impact on travel times. Higher hydraulic conductivity reduced travel time, while the influence of stream slope was more pronounced at higher slope levels. An increase in L_y shortened travel times, whereas an increase in L_x increased them. Additionally, the initial positions of the particles and their distances to the stream were found to have a significant impact on the model's performance in predicting travel times. The model’s performance was evaluated using error metrics such as the coefficient of determination (R²), mean absolute error (MAE), and mean absolute percentage error (MAPE), achieving high accuracy. The findings indicate that particle travel times to the stream can be effectively predicted using the XGBoost model. This study provides a practical and efficient model that can be utilized for managing stream-aquifer systems and analyzing pollutant transport. The results contribute to the determination of residence time dynamics in stream-aquifer interactions and provide a foundation for future studies.