Estimating seepage in heterogeneous earthfill dams on permeable foundations using explainable machine learning.

Predicting seepage discharge through heterogeneous earthfill dams founded on permeable foundations is essential for dam safety and long-term water-resources sustainability. Accordingly, this study evaluated five machine learning (ML) models viz Decision Tree (DT), Random Forest (RF), Stochastic Gradient Boosting (SGB), Light Gradient Boosting (LGB), and Categorical Gradient Boosting (CGB), to estimate seepage discharge using seven geometric and hydraulic input variables. The dataset was partitioned into training (80%), validation (10%), and testing (10%) subsets. To improve predictive capability, Bayesian Optimization (BO) was applied for hyperparameter tuning. Subsequently, model performance was examined using multiple error metrics, predicted–actual scatter plots, SHapley Additive exPlanations (SHAP) for interpretability, and k-fold cross-validation; finally, a rank-based analysis was used to consolidate the results across evaluation criteria. Overall, the tuned models exhibited substantial performance gains. In particular, the boosting-based methods consistently outperformed the DT and RF baselines. Among all candidates, the CGB model achieved the best overall performance, delivering near-perfect accuracy (R² = 0.9981 on validation and R² = 0.996 on testing). Moreover, cross-validation confirmed its robustness, as it produced the lowest RMSE values across the 10 folds. SHAP-based analysis further indicated that reservoir water depth and the core-to-shell hydraulic conductivity ratio are the dominant drivers of seepage behavior, whereas dam crest width and foundation depth exert secondary influence. To support practical use, a standalone desktop GUI was also developed to enable instant predictions with flexible input formats, batch evaluation, and reproducible export functions. Collectively, these results demonstrate that the CGB model provides a reliable and deployable framework for seepage prediction, maintaining prediction differences below 10% relative to prior numerical and empirical references across varying water depths.