Multi-fidelity Machine Learning of Planing Hulls Maneuvering and Seakeeping Dynamics
Abstract
The proposed effort fits into the Multi-disciplinary Multi-fidelity Simulation Drive Digital Design development tools that ONR is pursuing with direct involvement of the end user NSWC-SCD.-The accurate prediction of dynamics of high-performance planing hulls in waves is key for ensuring safety of personnel on board optimizing the sizing and design of the structural elements and improving their operational capabilities in high sea states. Recent ONR funded research contributed to advance the capabilities of Computational Fluid Dynamics models (namely Unsteady RANS) to predict-hydrodynamic loads (slamming) and motion response on small fast planning crafts in waves. These high-fidelity tools, though, require considerable computational resources and calculation time that make their systematic use prohibitive for design purposes, where all different operating scenarios, including different several sea states spectra, multiple relative incidence angles and several boat speeds are to be considered. In this new effort, we leverage results, methods and tools developed in previous research efforts for the high-fidelity prediction of highperformance high-speed planing hulls dynamics in wave. The completely new approach features Physics Informed Neural Networks-(PINNs) and Recursive Neural Networks (RNN), like DeepOnets, as a tool to create high-fidelity surrogate models of the system response to wave inputs or even to identify the mathematical structure of the dynamical system, i.e. the non-linear partial differential equations that describe the motion response of the hull in waves. We will explore possibilities to combine heterogeneous information,-such as Partial Differential Equation (PDE) modelling the physics, multi-fidelity data obtained from a minimal set of unsteady highfidelity CFD simulations, lower-fidelity motion prediction data, and eventually also experimental measurements. Differently from most fundamental research projects fusing on notional hull forms (often being far too simplified and of low practical-interest), in this proposal we specifically target methods and applications relevant to high-performance planing hulls of contemporary interest due to their superior hydrodynamic performance both in calm water and in waves, such as planing hulls with swept steps and the foil-assisted crafts.-High quality experimental tests will be carried out at the large towing tank of the US Naval Academy and used for validation as well as an alternative source of data for training the DNN. A combination of towing tank tests with forced (captive) motion in calm water and free response in waves will be designed and carried out using the fast-craft 3 degrees of freedom (pitch, heave, roll) towing rig and the horizontal planar motion mechanism (yaw and sway) in the USNA large towing tank. Two new models with significant geometric differences from the parent twin-step model will be designed, built and tested, for this new effort.-The fundamental open research questions that the project aims at answering are:-- Can Deep Neural Networks (DNN) be trained and used as surrogate model to predict planing hulls dynamics?-- Can DNN opportunely extended and trained predict the response of any hull belonging to a family of hull forms?-- Can DNN trained on a limited set of data predict the dynamics of the hull in any condition (e.g. any sea state)?-- What is the Accuracy and Efficiency (speed-up) of using DNN instead of direct CFD simulations-- Can the DNN be trained on multi-fidelity data, acquired from heterogeneous sources?-- Can the types of DNN used to predict the hull dynamics be extended and trained to predict the maximum loads in waves
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jun 13, 2024
- Source ID
- N000142412342
Entities
People
- Stefano Brizzolara
Organizations
- Office of Naval Research
- United States Navy
- Virginia Tech