A neural-corrector method for prediction of the vertical motions of a high-speed craft

Abstract

A machine learning method is developed which is capable of making predictions of nonlinear ship motions in a range of wave conditions when trained on response data from only a single seaway. The method is formulated around the equations of motion in the time domain, but the equations are augmented with data-driven terms that act to correct the force-balance in the equations. The data-driven nonlinear forcing terms are modeled using Long Short-Term Memory (LSTM) recurrent neural networks. The resulting hybrid governing equations are solved numerically. Predictions from the method are compared to nonlinear test data of 2-DOF motion of a Generic Prismatic Planing Hull (GPPH) at forward speed in head seas, with time histories given for both regular and irregular waves. The training data requirements to classify a specific seaway are investigated and quantified. Predictions over a range of significant wave heights and peak periods are performed using training data from only a single seaway to show the effectiveness of the method in generalizing across different environmental conditions.