Hybrid FEM and machine learning framework for probabilistic seepage failure prediction in embankment dams under drawdown conditions

Conventional embankment dam safety assessments under reservoir drawdown conditions are typically based on deterministic finite element and limit equilibrium analyses. That use single representative values of hydraulic and shear strength parameters. Such approaches often neglect the inherent uncertainty and spatial variability of geomaterials leading to potential inaccuracies in predicting seepage behavior, slope stability and failure risk. Although probabilistic methods can better represent these uncertainties, their computational complexity limits their application in routine dam safety evaluations. Furthermore, the integration of physical based numerical modeling with machine learning for probabilistic seepage failure assessment remains insufficiently explored. To address these limitations, this study proposes a hybrid deterministic probabilistic framework that integrates transient seepage and stability analyses with probabilistic risk assessment and explainable machine learning. The framework was applied to the Ribb Embankment Dam in Ethiopia under steady state, rapid drawdown (RDD) and slow drawdown (SDD) conditions. Transient hydraulic and stability responses were simulated using SEEP/W and SLOPE/W based on Richards’ equation and the Morgenstern Price method. The generated hydraulic and stability indicators were then incorporated into a probabilistic failure model and used to train Random Forest (RF), XGBoost and Support Vector Machine (SVM) models. Probabilistic analysis indicated that 85.75% of the simulated cases were associated with moderate failure probability range of 0.10 ≤ Pf < 0.30. Also 13.75% with high failure probability of 0.30 ≤ Pf < 0.50 and only 0.50% exceeded the critical threshold of Pf ≥ 0.50. The machine learning models achieved excellent predictive performance with XGBoost attaining R² = 0.9998 and RMSE = 0.0012. While Random Forest provided the best testing performance with R² = 0.9986 and RMSE = 0.0034. All models achieved ROC–AUC values of 1.000 and Brier scores between 0.048 and 0.051 that indicating excellent discrimination and calibration. SHAP analysis identified water (Hydraulic) gradient and water pressure as the most influential factors controlling seepage failure probability. The main contribution of this study is the development of a hybrid framework that combines physical based numerical modeling, probabilistic risk assessment and explainable machine learning for seepage induced instability evaluation. The proposed approach provides accurate failure probability prediction, improved model interpretability and a practical basis for risk informed embankment dam safety management and early warning systems.