Adaptive Multihazard Modeling Predicts Rainfall-Driven Dam Failure: a case

The safety of Embankment dams during the Construction phase has to be ensured accurately and it requires a strong integration between physical principles with data-driven forecasting. This study develops an adaptive multi-hazard modeling framework that hybridizes the Finite Element Method (FEM) with a deep learning model (ANN-LSTM) to predict rainfall-induced seepage failures in embankment dams. The model's performance was tested and enhanced using the Megech Dam case study. After being trained using FEM-based PCA clustered data, the model was evaluated against an alternative two-year (104-week) monitoring dataset. This information was important since it documented the dam's growing instability and final failure during construction. Results demonstrate that the hybrid framework achieves high predictive accuracy (MAE = 0.027, R² = 0.94) and identifies a critical 2–3 week lag between rainfall peaks and failure-triggering pore pressure buildup. An integrated attention mechanism independently pinpointed antecedent sequences, highlighting weeks 25–30 and 75–80 as high-risk periods. The model has significantly enhanced the standalone FEM (MAE = 0.098) and non FEM based data driven approaches, which learned nonexistent correlations. This work can be taken as a tool which significantly contributed for early warning of failures and moves dam safety management from a reactive approach.