Machine Learning Prediction of Dam-Break Surge Dynamics Trained on Physics-Based Simulations

Abstract

Accurate and rapid prediction of dam-break flows is critical for effective flood management and emergency response. This study presents a one-dimensional, physics-based numerical model of dam-break flows in prismatic channels, along with machine learning models trained on the outputs of the physics-based model. The numerical model solves the Saint-Venant equations using the MacCormack finite difference scheme. It is validated against experimental data and analytical solutions and is used to assess the influence of channel cross section, bed slope, and roughness on key surge characteristics, including surge height, velocity, and wavefront celerity. Two machine learning models are developed: a multilayer perceptron (MLP) and support vector regression (SVR). The MLP model achieves higher accuracy, with R2 values of 0.99, 0.98, and 0.96 for nondimensional surge height, velocity, and wavefront celerity, respectively. The machine learning models are tested for a real-world dam-break case and are further integrated with an analytical equation to predict attenuation of peak discharge, demonstrating their applicability for real-world flood forecasting and emergency planning.