Extreme Value Problems in Naval Hydrodynamics
Abstract
Extreme response of naval platforms at sea is a long-standing problem that still bothers the Navy today. Recent novel designs of Littoral Combat Ships, Expeditionary Fast Transports and DDG destroyers all turned out to be vulnerable to some extent to certain waveor flow conditions. These extreme conditions are not encountered in the design process mainly because of two reasons: (1) Reduced-order models such as large eddy simulations (LES) and Reynolds-averaged Navier-Stokes (RANS) simulations involved in platform designsare usually not capable of reliably predicting the extreme events. (2) Even with an accurate model available, computing the occurrence probability of the extreme events in a random ocean environment is extremely difficult since seemingly infinite number of simulations need to be conducted to capture the probability. These two potential flaws in design will be more detrimental for next-generation autonomous platforms, which will undergo radical change of the hull form(due to the need to load new weapons and equipment) and likely must undertake tasks in more hostile sea environment. In the proposed project, we will solve two classes of fundamental problems. The class-1 problem is to endow a parameterized reduced-order model the capability to deterministically predict extreme events that are supposed to be predicted well only by the full model. The class-2 problem is to evaluate the extreme-event probability for a platform (or other naval equipment) subject to random ocean loading (from wave or flow), given a model that computes the platform response accurately but with a high computational cost.Both classes of problems will be solved for a major example case, and then extended to other cases. For class-1 problem, we will consider the construction of a wall model (the wall shear stress in a parameterized form) in the wall-modeled LES (WMLES), to predict the near-wall extreme pressure in channel flow. We will approach this problem using multi-agent reinforcement learning (MARL), with multiple agents distributed on the surface of the body. Each agent learns a policy to assign instantaneous wall shear stress according to the surrounding flow state, with the policy function trained with a reward that puts more weight on the accurate prediction of large near-wall pressure (compared to DNS data). For the example case in class-2 problem, we will consider the quantification of extreme ship motion probability in irregular ocean waves. The key in solving this problem lies in the parameterization of random irregular waves into a relatively low-dimensional parameter space, and application of sequential sampling method to minimize the number of response evaluations needed to evaluate the extreme response probability. The wave information lost in the parameterization step (due to dimension reduction) will be recovered through a novel approach applying heteroscedastic variance model in establishing the stochastic response function. Approved forpublic release.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 15, 2024
- Source ID
- N000142412266
Entities
People
- Yulin Pan
Organizations
- Board of Regents of the University of Michigan
- Office of Naval Research
- United States Navy