Date of Award
1-1-2025
Document Type
Dissertation
Degree Name
Ph.D. in Engineering Science
First Advisor
Jacob Najjar
Second Advisor
Ahmed Al-Ostaz
Third Advisor
Hakan Yasarer
School
University of Mississippi
Relational Format
dissertation/thesis
Abstract
This dissertation presents a multi-faceted framework for the prediction of geotechnical and hydraulic failure mechanisms in earthen structures, with a focus on levees and embankments. The study integrates laboratory testing, numerical modeling, and artificial intelligence (AI) to enhance the assessment and forecasting of levee performance under static and dynamic loading conditions.
The first component of the research applies various machine learning (ML) algorithms to predict Unconfined Compressive Strength (UCS) of soils using basic geotechnical parameters. Experimental data from lab-tested soil mixtures were used to train and compare models including Artificial Neural Networks (ANN), Gradient Boosting, Random Forests, K-Nearest Neighbors, Support Vector Machines, and traditional regressions. Among all techniques, ANN yielded the highest prediction accuracy and was embedded into an Excel-based graphical user interface (GUI) for practical applications.
The second phase involved advanced numerical modeling using FLAC3D to simulate a range of levee failure scenarios. A simulation matrix of 516 cases was developed to analyze the effects of varying levee geometry and soil properties under static and pseudo-static (seismic) loading. The models captured shear failure, piping, and bearing capacity failure in both the embankment and foundation. Two ANN models were trained using the simulation results—one to predict the factor of safety (FOS), and another to estimate failure indices—demonstrating high predictive performance and robustness. These models were implemented into a user-friendly GUI to support risk evaluation and design.
The final study utilized 5,000 breach simulations from DLBreach, a physics-based hydraulic software, to train ANN models that predict key hydraulic outcomes, including peak outflow discharge and time to peak. The inputs considered hydraulic and erodibility parameters such as critical shear strength, erosion coefficient, headcut coefficient, and soil characteristics. The best-performing ANN was integrated into another GUI, enabling efficient prediction of breach behavior.
Together, these studies showcase the potential of machine learning models, trained on both experimental and simulation data, to significantly enhance the prediction of levee performance and breach consequences. The developed tools can support decision-making in flood risk management, emergency planning, and geotechnical design.
Recommended Citation
Alqudah, Mudhaffer, "Levee Breach Analysis Using Numerical Simulation Aided by Machine Learning" (2025). Electronic Theses and Dissertations. 3227.
https://egrove.olemiss.edu/etd/3227