Degree
Doctor of Philosophy (PhD)
Department
Civil & Environmental Engineering
Document Type
Dissertation
Abstract
This dissertation proposes the best usage practices and development of state-of-the-art machine learning and deep learning models to predict local flow features and scour around obstacles. The motivation arises from the necessity to accurately model sediment transport with reasonable computational expense as well as data compression for efficient data storage. The development of such models will aid in understanding and predicting erosion, deposition, and morphological changes in various aquatic environments. Traditionally, scour and erosion estimations rely on empirical models, without necessarily considering the momentum exchange between sediment and fluid phases as well as inter-particle interactions. Computational fluid dynamics (CFD) models, in this regard, become helpful but are computationally expensive for most of the applications. This study assessed the suitability of advanced data-driven techniques — Proper Orthogonal Decomposition with Long Short-Term Memory networks (POD-LSTM), $\beta$ - Variational Autoencoders with LSTM ($\beta$-VAE - LSTM), Fourier Neural Operators (FNOs) in emulating sediment transport simulations around a submerged cylinder as well as Generative Adversarial Networks (GANs) to reconstruct the whole spatiotemporal data when some snapshots are missing. The results show that $\beta$-VAE - LSTM can predict flow and scour patterns with high accuracy and efficiency, offering a promising alternative to traditional CFD methods. Building on this, neural operator frameworks such as FNOs were investigated to predict the spatiotemporal evolution of local scour. A hybrid U-FNO architecture, which combines global spectral convolutions with U-Net refinement pathways, achieved stable long-horizon rollouts across all flow variables while providing orders-of-magnitude computational speedup compared to SedFoam simulations. The GAN framework is investigated for reconstructing velocity and vorticity fields beneath a submerged cylinder. Although less effective for long-term forecasting, it shows strong capability in gap-filling and short-horizon predictions.
Date
11-1-2025
Recommended Citation
Nazari, Haq Murad, "Reduced-order and Generative Modeling in Fluid Mechanics and Sediment Transport to Predict Local Flow and Scour Predictions around Obstacles" (2025). LSU Doctoral Dissertations. 6940.
https://repository.lsu.edu/gradschool_dissertations/6940
Committee Chair
Dr. Celalettin E. Ozdemir