Authors: Shivakumar Balachandran, Ezzat Elalfy, Mahdi Erfani, Jani Fathima Jamal, Jasim Imran, Hanif Chaudhry, Erfan Goharian – University of South Carolina
Presentation Type: Lightning Talk and Poster
Title: Machine Learning Models for Levee Breach Flow Estimation
Abstract: This study represents advancement in the application of Machine Learning (ML) models as surrogates for hydraulic modeling, with a focus on estimating levee breach discharge and approach depth in rectangular channels. A Multi-Layer Perceptron (MLP) model is trained by integrating results from a 2D numerical model of levee breaches. The accuracy of the 2D numerical model is validated using previous experimental setups and tests at the University of South Carolina. The input variables for the surrogate models include inlet Froude number, outflow Froude number, relative breach width, and relative levee height, all non-dimensionalized to make the model scale-invariant, enabling its applicability across various scenarios. The MLP achieved R-squared values of 0.99 for both breach discharge and approach depth. The model exhibited capabilities with a Root Mean Squared Error (RMSE) of 0.029 and 0.015 for relative breach discharge and approach depth, respectively, while extrapolating beyond the training data. Furthermore, the MLP model demonstrated performance with an RMSE of 0.064 and 0.0138 when tested using experimental data for relative breach discharge and approach depth, respectively. Future research will extend this study to incorporate the effect of channel cross-sectional parameters on breach outflow and approach depths.