A Computational Fluid Dynamics Driven Machine-Learning Framework for Observation and Quantification of Extreme Ship Reponses

Abstract

The prediction of extreme ship responses remains an important and longstanding topic in ship hydrodynamics, with continued focus on developing probabilistic methods based on simplified descriptions of the hydrodynamics that mainly produce qualitative observations. While simpler hydrodynamic formulations provide insight into extreme events, their underlying assumptions can prevent accurate quantitative representations of the mechanisms ultimately responsible for the extreme responses. With this limitation in mind, research must strive to utilize increasingly more accurate hydrodynamic formulations to provide quantitative observations and statistical characterization of extreme ship response events. Additionally, previous work in the prediction and quantification of extreme events has involved simplifications such as only considering zero speed or constant speed and heading. Consequently, resulting calculations have largely neglected free-running vessels traveling with 6 degrees-of-freedom (6-DoF), where surge, sway, and yaw motions, in conjunction with propeller and rudder forces, contribute to extreme events and failures. The following research develops the CCS extreme event probabilistic framework capable of both observing and quantifying the probability of extreme events by integrating the critical wave groups (CWG) extreme event probabilistic method, fully nonlinear Computational Fluid Dynamics (CFD) to achieve a high-fidelity representation of the hydrodynamics, and long short-term memory (LSTM) neural networks to build surrogate models of the CFD predictions to improve the overall computational efficiency. This framework intentionally accommodates free-running vessels as well as various physical dynamical mechanisms that could lead to extreme events, without the need for intrusive dynamic constraints. As a result, this approach provides an avenue forward for high-fidelity extreme event analysis at a practical computational cost.

This dissertation demonstrates the CCS framework on case studies utilizing a two-dimensional (2-D) midship section and a three-dimensional (3-D) representation of the Office of Naval Research Tumblehome (ONRT) hull form. The CCS framework is first demonstrated with a 2-D midship section of the ONRT in Sea State 7 beam seas that is only free to heave and roll. The case study implements the CCS framework with and without the LSTM neural networks to understand both the accuracy of the framework as well as the accuracy of the surrogate modeling technique for extremes. The CCS framework is able to produce responses and probability predictions that are representative of a purely CFD-driven CCS framework with 200 high-fidelity CFD training simulations, corresponding to seven orders of magnitude of computational cost savings when compared to a Monte Carlo approach. The other case study implements the CCS framework for a free-running 3-D ONRT, traveling in stern-quartering Sea State 7 seas and free to move in all 6-DoF. The case study tests the ability of the CCS framework to handle arbitrary frames of encounter when enforcing initial conditions as well as the ability of the LSTM neural networks to represent the extreme 6-DoF vessel response. Similarly to the 2-D case study, the probability of exceedance calculations from the surrogate models converge at around 200 training runs and produce LSTM predictions that are representative of the CFD validation simulations. The 3-D case study provides a total of three orders of magnitude in computational cost savings compared to Monte Carlo.

This dissertation develops the CCS framework and an LSTM neural network surrogate modeling methodology, as well as showcases significant advancement in the observation and probabilistic quantification of extreme ship response events with significant reductions in computational cost.