Assessing the impact of groundwater abstraction and concrete dam fractures on saltwater intrusion using numerical modeling and interpretable machine learning.

Reliable prediction of SWI is essential for protecting coastal groundwater. In this study, the SWI wedge length in a sloping coastal aquifer controlled by groundwater abstraction and a fractured underground dam is estimated. To achieve this, a dataset consisting of eight dimensionless inputs, derived from prior SEAWAT numerical scenarios, was used to train six machine learning models (linear, nonlinear, and ensemble) to predict the relative SWI wedge length (L/H). First, the dataset underwent thorough examination using hypothesis testing, multicollinearity analysis, and correlation analysis to assess the significance of predictors, their interdependencies, and their relationships with L/H. Specifically, statistical tests, including ANOVA and Z-tests, were employed to identify the key predictors. Furthermore, multicollinearity was assessed using the Variance Inflation Factor (VIF), which helped identify potential redundancies. Subsequently, a cosine amplitude sensitivity analysis was used to quantify the relative influence of each input on L/H. In addition, Bayesian optimization was applied to fine-tune the hyperparameters of each model for optimal performance. The performance of the models was then evaluated using 10-fold cross-validation, regression metrics, and external validation on independent numerical scenarios from the Akrotiri coastal aquifer (Zakaki, Cyprus). Additionally, explainable ML techniques viz Shapley-Additive-Explanations (SHAP) and Partial-Dependence-Plots (PDP), were used to interpret model behavior. Results showed that the ensemble models outperformed alternatives. Among them, the XGB model provided the most consistent accuracy during the testing stage, yielding R