Armanuos et al. (2026) Assessing the impact of groundwater abstraction and concrete dam fractures on saltwater intrusion using numerical modeling and interpretable machine learning

Identification

• Journal: Scientific Reports
• Year: 2026
• Date: 2026-03-13
• Authors: Asaad M. Armanuos, Martina Zeleňáková, Mohamed Kamel Elshaarawy
• DOI: 10.1038/s41598-025-27998-4

Research Groups

• Irrigation and Hydraulics Engineering Department, Faculty of Engineering, Tanta University, Tanta, Egypt
• Institute of Environmental Engineering, Faculty of Civil Engineering, Technical University of Košice, Košice, Slovakia
• Civil Engineering Department, Faculty of Engineering, Horus University-Egypt, New Damietta, Egypt

Short Summary

This study develops and validates machine learning models to predict the relative saltwater intrusion (SWI) wedge length (L/H) in coastal aquifers, considering groundwater abstraction and fractured underground dams. The XGBoost model demonstrated superior accuracy (R²=0.9978, RMSE=0.216) and identified the relative recharge well rate as the dominant predictor, offering a robust tool for SWI management.

Objective

• To assess the impact of groundwater abstraction and fractured concrete subsurface dams on saltwater intrusion (SWI) wedge length (L/H) in coastal aquifers.
• To develop and validate interpretable machine learning models (MLR, SVR, GPR, DT, RF, XGB) for predicting the relative SWI wedge length (L/H) using dimensionless parameters derived from numerical simulations.
• To identify the most influential parameters affecting SWI dynamics and provide practical, user-friendly tools for groundwater management.

Study Configuration

• Spatial Scale: Two-dimensional cross-section of a sloping coastal aquifer; model domain of 3000 meters in length and 100 meters in depth; cell size of 6 meters by 6 meters. Case study: Akrotiri coastal aquifer, Cyprus, with aquifer thickness varying from 10 meters to over 100 meters.
• Temporal Scale: Steady-state conditions for numerical simulations and model validation.

Methodology and Data

• Models used:
  • Numerical: SEAWAT (for generating the initial dataset and independent validation).
  • Machine Learning: Multiple Linear Regression (MLR), Support Vector Regression (SVR), Gaussian Process Regression (GPR), Decision Tree (DT), Random Forest (RF), Extreme Gradient Boosting (XGB).
  • Interpretability: SHAP (Shapley Additive Explanations), PDP (Partial Dependence Plots).
• Data sources:
  • A dataset of 438 numerical scenarios from a prior study (Armanuos et al., 2022) for training and testing ML models.
  • Independent numerical scenarios (35 runs) from the Akrotiri coastal aquifer (Zakaki, Cyprus) simulated with SEAWAT for external validation.
  • Input parameters were dimensionless ratios: relative density (ρf / ρs), relative fracture aperture height (Hf / H), relative fracture aperture diameter (Df / H), relative subsurface dam height (Hd / H), relative subsurface dam distance (Ld / H), relative well height (Hw / H), relative well distance (Lw / H), and relative well abstraction rate (Q_w / KH²).

Main Results

• Ensemble machine learning models, particularly XGBoost, demonstrated superior predictive performance compared to linear, non-linear, and single-tree models.
• The XGBoost model achieved excellent accuracy on the testing set (R² = 0.9978, Root Mean Squared Error (RMSE) = 0.216, Mean Absolute Error (MAE) = 0.058, Mean Absolute Relative Error (MARE) = 0.098) and maintained strong training performance (RMSE = 0.037).
• Independent validation against 35 SEAWAT numerical scenarios for the Akrotiri coastal aquifer confirmed high fidelity and generalizability (R² = 0.997, RMSE = 0.157).
• SHAP and PDP analyses revealed that the relative recharge well rate (Qw / KH²) was the most dominant predictor of the relative SWI wedge length (L/H), followed by relative fracture height (Hf / H), relative fracture diameter (Df / H), and relative well distance (Lw / H).
• User-friendly desktop and web-based graphical user interfaces (GUIs) were developed for rapid and accessible prediction of L/H.
• Parametric analysis showed SWI wedge length is highly sensitive to abstraction well rate, abstraction well location, dam depth, and fracture diameter, and moderately sensitive to dam location, fracture height, and abstraction well height.

Contributions

• Developed a novel, comprehensive approach integrating numerical modeling, dimensional analysis, and interpretable machine learning to predict dynamic saltwater intrusion (SWI) wedge length (L/H) in coastal aquifers.
• Addressed a significant research gap by focusing on the dynamic L/H ratio and the combined effects of groundwater abstraction and fractured subsurface dams, which were largely overlooked in previous static SWI prediction studies.
• Demonstrated the superior performance and robustness of ensemble machine learning models, particularly XGBoost, for accurate and reliable SWI prediction in complex hydrogeological settings.
• Enhanced model transparency and trust through the application of explainable AI techniques (SHAP and PDP), providing critical insights into the influence of various hydrogeological and management parameters.
• Provided practical, accessible tools (desktop and web GUIs) for groundwater managers and stakeholders to facilitate informed decision-making and effective SWI control in coastal regions.

Funding

Not specified in the paper.

Citation

@article{Armanuos2026Assessing,
  author = {Armanuos, Asaad M. and Zeleňáková, Martina and Elshaarawy, Mohamed Kamel},
  title = {Assessing the impact of groundwater abstraction and concrete dam fractures on saltwater intrusion using numerical modeling and interpretable machine learning},
  journal = {Scientific Reports},
  year = {2026},
  doi = {10.1038/s41598-025-27998-4},
  url = {https://doi.org/10.1038/s41598-025-27998-4}
}