Optimizing Ship Airwake Database for Predicting Autospectra Using Deep Learning

Abstract

Airwakes are shed behind a ship's superstructure, and they represent a highly turbulent and rapidly distorting flow field. Because of this complexity, they require a baseline: A database of flow velocity points, preferably through experiments and also through high-fidelity computational fluid dynamics simulations. However the database is prohibitively costly, and in situ measurements, at a constant wind angle, are often affected by the weather. A previously developed perturbation scheme helps to generate the autospectrum in closed form from a given wind database, specifically, from the numerically extracted autospectrum for this database. The primary objective of this paper is to demonstrate how deep neural networks can be fruitfully exploited to reduce the dependence on a database of flow velocity points in generating the autospectra, by reducing the amount of in situ data required to achieve the same accuracy as established methods. Two machine learning approaches are developed to predict the autospectrum for a data set when provided with a subset of the wind data set at the input and are assessed using cross-validation. In the first, a neural network is applied to predict the perturbation scheme. Though this method did not predict the autospectrum better than the previously developed perturbation scheme, it laid the foundation for the second method. In the second, a neural network is applied to the autospectra which can be numerically extracted from a database. An improvement is found in the average error computed over all data sets for the model-predicted autospectrum compared to the average error in the autospectrum predicted using the conventional method. The model provided better results for 51.4% of the data sets compared to the conventional method, with many data sets showing marked improvement. The database measured from the Naval Academy YP676 training vessel is used for training and testing the neural network.