A Hybrid Investigation Combining Numerical and Experimental Models with Machine Learning Techniques to Study the Erosion Rate and Peak Outflow for Earth-Fill Dam Breaches

Abstract

Understanding and accurately predicting the outflow hydrograph from embankment dam breaches is essential for managing the associated flood hazard and improving emergency preparedness. This work simulates the breaching process using a high-resolution 3D computational fluid dynamics (CFD) model, a critical natural hazard for earth-fill dams under overtopping conditions. The model was validated against the experimental data, showing high accuracy in predicting breach development and failure timing. A parametric analysis was performed to assess the influence of the initial breach geometry on erosion dynamics. The results indicated a high sensitivity, while increasing the breach width by 5% led to an average 11% increase in the erosion rate, and decreasing the depth by 5% caused an average 16.5% rise. To enhance predictive capabilities for this hazard, a multilayer neural network (MLNN) was trained on the CFD-generated dataset. The network utilized breach geometry and time as inputs to forecast the peak outflow and erosion rate, achieving excellent accuracy (RMSE = 0.019, R2 = 0.99). This integrated modeling strategy combines data-driven learning with physics-based simulation and demonstrates its effectiveness for laboratory-scale dam breach modeling. This approach is a step toward more efficient surrogate-based tools for flood risk analysis, though its extension to full-scale dams and varied material properties requires additional validation and scaling analyses beyond the scope of this work.