3D wave simulation based on a deep learning model for spatiotemporal prediction. (1st November 2022)

• Record Type: Journal Article
• Authors: Li, Ying; Zhang, Xiaohui; Cheng, Lingxiao; Xie, Ming; Cao, Kai
• Is Part Of: Ocean engineering. Volume 263(2022)
• Publication Date: 2022-11-01
• Subjects: Real-time wave simulation -- ConvLSTM -- FFT -- Wave spectrum -- Deep learning; Ocean engineering -- Periodicals; Ocean engineering; Periodicals; 620.4162
• DOI: 10.1016/j.oceaneng.2022.112420
• Languages: English
• ISSNs: 0029-8018

Abstract

Abstract: Ocean wave simulations must be conducted in real-time and are more complicated than other natural scenery simulations. This study proposes a novel ocean wave simulation method that inputs the spatiotemporal sequences of the wave height field obtained by a wave spectrum formula and the fast Fourier transform (FFT) algorithm into a convolutional long short-term memory (ConvLSTM) training model. The method resolves the problems of poor real-time performance and authenticity in the traditional ocean wave simulation process. The ocean wave simulation method calculates the wave height field rapidly using the ConvLSTM-based model rather than the traditional FFT method. Finally, it accelerates the wave simulation process and predicts the height field at a future time. The model was evaluated in a simulation experiment on two kinds of wave spectra. The experimental results confirmed the realism of the waves simulated by the proposed model. The computational speed of the ConvLSTM model exceeds that of the FFT method, especially as the sample size and length of the prediction sequence increase, indicating the effectiveness and feasibility of the ConvLSTM model in accelerating the FFT algorithm. Graphical abstract: Image 1 Highlights: The wave height fields obtained by FFT algorithm can be regarded as a sequence of height maps with spatiotemporal features. A deep learning model for spatiotemporal prediction named ConvLSTM can replace FFT algorithm in predicting height fields. The ConvLSTM model can output one or more height maps in a single execution and outperforms FFT in real-time performance.