An Adapted ResNet-50 Architecture for Predicting Flow Fields of an Underwater Vehicle

Abstract

This study introduces FlowRes, an adapted ResNet-50 architecture, to predict flow fields around underwater vehicles, aiming to improve the efficiency of Computational Fluid Dynamics (CFD) through deep learning. The background highlights the necessity for rapid and accurate flow field predictions to enhance the hydrodynamic shape design of an Autonomous Remote-Controlled Vehicle (ARV) for inspection of offshore energy underwater infrastructure. Employing a decoder-only CNN-based model, the methodology involves modifying ResNet-50 for image-to-image generation, focusing on flow field visualization of underwater vehicles using a compact dataset. Results from training the model with 240 paired entries of flow fields and design parameters show significant computational speed improvements, with generated images deviating less than 1% from traditional CFD-generated images. The conclusions underline the potential of integrating advanced deep learning techniques with CFD, demonstrating FlowRes’s capability in significantly accelerating the design process of underwater vehicles by offering a faster, more resource-efficient alternative to traditional methods.

Keywords

• Computational fluid dynamics
• deep learning
• flow field prediction
• underwater vehicles design