Data-Driven Identification of Critical Wave Groups

Abstract

. Accurate and efficient prediction of extreme ship responses continues to be an important and challenging problem in ship hydrodynamics. Probabilistic frameworks in con-junction with computationally efficient numerical hydrodynamic tools such as volume-based and potential flow methods have been developed that allow researchers and ship designers to better understand extreme events. However, the ability of these tools to represent the physics quantitatively during extreme events is limited and not robust to different problems. There-fore, model testing will continue to be important in analysis, and more emphasis will be placed on high fidelity computational fluid dynamics (CFD) simulations. Experiments and CFD both come at well documented costs and require a systematic approach to target extreme events. The critical wave groups method (CWG) has been implemented with CFD, and the integra-tion of high fidelity simulations with extreme event probabilistic methods has been previously showcased. The implementation of CWG with CFD is achieved by embedding deterministic wave groups into previously run irregular wave trains such that the motion state of the ship at the moment of encountering the wave group is known. Embedding the deterministic wave groups into an irregular wave train results in a composite wave train that can be evaluated with numerical hydrodynamic simulation tools such as CFD, or even a model test. Though the CWG method does allow for less simulation time than a Monte Carlo type approach, the large number of runs required may still be cost-prohibitive. The objective of the present work is to develop an approach where a limited set of expensive simulations or experiments build a time-accurate long short-term memory (LSTM) neural network model that rapidly identifies critical wave groups that lead to a response exceeding a specified threshold. This paper compares the LSTM modeling approach of building a single neural network for all wave groups to establishing an ensemble of neural networks, each responsible for wave groups with specific parameters. The ensemble approach showcases better accuracy, a higher convergence with respect to data quantity, and produces responses that are representative of the CFD simulations.