A novel Bayesian deep learning method for fast wake field prediction of the DARPA SUBOFF

• Published in: Applied ocean research Vol. 150; p. 104074
• Main Authors: Xie, Xinyu, Zhao, Pu, Bian, Chao, Xia, Linsheng, Ding, Jiaqi, Wang, Xiaofang, Liu, Haitao
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.09.2024
• Subjects: Bayesian deep learning; Convolutional neural network; Paddle disk surface; SUBOFF; Uncertainty quantification; Wake field; SUBOFF; Uncertainty quantification; Bayesian deep learning; Paddle disk surface; Convolutional neural network; Wake field
• ISSN: 0141-1187
• DOI: 10.1016/j.apor.2024.104074

The accurate and rapid prediction of wake flow characteristics is of great significance for the design of underwater vehicles. This paper develops a data driven Bayesian deep learning framework, named WakeNet, to accurately and quickly predict the wake velocity field on the paddle disk surface of the DARPA SUBOFF model under varying geometry configurations. Specifically, (i) the proposed WakeNet presents a coordinate affine transformation technique to address the issue of handling unstructured wake flow data with varying geometries; (ii) it develops a residual modeling strategy built upon a modern version of the U-net architecture to capture the multi-scale wake flow characteristics; and finally, (iii) it exploits the MC dropout method to achieve uncertainty quantification of the prediction, and builds the uncertainty based active learning framework to effectively update the model. Numerical experiments have verified the capability of the proposed WakeNet model for the accurate prediction and uncertainty quantification of the wake flow field with varying geometries. Besides, the proposed model yields a speed up factor of 7000× compared to the conventional numerical simulation solver. •A novel Bayesian deep learning method for fast prediction of the wake field of DARPA SUBOFF.•A coordinate affine transformation method for tackling the unstructured wake flow data under varying geometries.•A residual modeling strategy built on the U-net architecture to capture the multi-scale wake flow characteristics.•A Monte Carlo (MC) dropout based uncertainty quantification and active learning framework have been built.•The proposed model achieves comparable yet 7000× faster wake field predictions in comparison to the CFD solver.