CFD-driven physics-informed generative adversarial networks for predicting AUV hydrodynamic performance

A novel framework for predicting the hydrodynamic performance of autonomous underwater vehicles (AUVs) is proposed based on computational fluid dynamics (CFD) and physics-informed generative adversarial networks (PIGAN). The physics information is embedded in the generator and discriminator, improving prediction accuracy and efficiency. The antenna position, height, and propeller diameter are modelled parametrically. Ultimately, 27 design schemes are constructed by combining the three variables and their three levels in full permutations. In the present study, 1440 groups of sample points and 270 groups of response points are obtained through CFD simulation for PIGAN training and testing. Furthermore, the prediction performance of CFD and PIGAN is compared, and the results show that PIGAN has significant advantages in efficiency and generalisability. The effects of antenna position and height on the AUV's flow field and motion performance are systematically analysed using the validated method. The hull-appendage-propeller interaction is calculated and analysed. The results indicate that the size and layout of the appendage significantly affect the turbulence intensity, especially at the top of the appendage. External devices should be prioritised in an embedded design, with the necessary external sensors in an optimised layout to provide a positive wake flow. • A novel method is proposed based on CFD and physics-informed generative adversarial networks. • The generator employs a convolutional neural network, and the discriminator adopts a multilayer perceptron. • The appendages have a significant effect on the turbulence intensity, especially in the top zone of the antenna. • The proposed CFD-driven PIGAN model has high efficiency, accuracy and robust generalisability.