Exploring Complex Problems in Fluid Dynamics: from CFD to Experiments Leveraging ML
Author(s)
del Aguila Ferrandis, Jose
DownloadThesis PDF (166.5Mb)
Advisor
Triantafyllou, Michael
Terms of use
Metadata
Show full item recordAbstract
Structures operating in the marine environment are subject to large steady and unsteady forces, typically at high Reynolds number. Given the current limitations on CFD at high Reynolds number, and the high expense to conduct experiments at relevant scales, design is constrained by the limited available data, incomplete knowledge of the principal physical mechanisms, and restricted parametric searches, ML offers opportunities to overcome these limitations.
In this thesis we demonstrate we apply ML methods to three engineering problems of high importance to theory and applications:
1. Optimizing Vortex Generators to Reduce Ship Form Drag: In the quest for reducing ship emissions, it is imperative that the fluid mechanics of ship resistance be explored for improving propulsive efficiency. Form drag is a significant part of the resistance of high block coefficient ships and remains a last frontier for hull ship optimization. Weexplore the optimization of vortex generators (VG) as a powerful tool for reducing f low separation.
2. Mapping the Properties of Fluid Forces in Vortex Induced Vibrations of SCR Risers: Vortices form in the wake of bluff bodies as a result of flow instabilities that are hard to study parametrically, especially for complex structures such as steel catenary risers (SCR). The resulting vibrations are of theoretical and practical importance. By using experimental and field data we can extract hydrodynamic databases the incorporate known physics, fill the parametric space, and provide new knowledge. By focusing on the SCR vibratiosn,w we demonstrate that we not only extract physics, but can provide accurate predictions as well.
3. Causal Learning of Large Amplitude Ship Motions with Emphasis on Parametric Rolling: Predicting ship motions in severe sea states is complex due to the nonlinear wave-body interactions involved. This section introduces a simulation approach utilizing neural networks trained on stochastic wave elevations from multiple sea states.
The trained networks can predict core vessel motions, including the challenging phenomenon of parametric rolling. Once trained using detailed CFD simulations, these networks provide swift and efficient vessel dynamics predictions. The research also explores the statistics of non-linear motions, aiming for consistent and accurate predictions across different wave conditions. This methodology, influenced by the universal approximation theorem for functionals, represents a significant advancement in addressing engineering challenges.
In summary, these studies emphasize the role of ML a instrumental tool in advancing marine systems, driving them toward increased efficiency and adaptability.
Date issued
2024-05Department
Massachusetts Institute of Technology. Department of Mechanical Engineering; Massachusetts Institute of Technology. Center for Computational Science and EngineeringPublisher
Massachusetts Institute of Technology