Explainable machine learning surrogate modeling for exit gradient prediction in hydraulic structures using numerical simulations

ABSTRACT Accurate prediction of exit gradients is essential for designing impounding hydraulic structures, such as dams and levees, to mitigate seepage-induced piping failures. Conventional analytical methods often neglect boundary effects and anisotropy, limiting their applicability in complex geological settings. This study develops an XGBoost-based machine learning surrogate model trained and tested on 8,000 MODFLOW numerical simulations to predict exit gradients under varying hydraulic and geological conditions. The data set covers numerous parameters, including anisotropy, head differences, structure width, cut-off wall depth, aquifer thickness, and uninterrupted riverbed length. The coefficient of determination for test data is 0.88, demonstrating reliable exit gradient prediction by the XGBoost model. To enhance model interpretability, we employ the SHAP (SHapley Additive exPlanations) framework, which identifies anisotropy and the ratio of the cut-off wall depth to aquifer thickness as the most significant factors influencing exit gradients. Our results show that anisotropy plays a dominant role when the cut-off wall is relatively shallow compared to the hydraulic structure width, while its impact diminishes with larger vertical anisotropy. These findings highlight the importance of constraining subsurface properties in hydraulic structure design and provide a comprehensive model-driven approach for seepage control strategies that could be applied to other complex engineering problems.