The emergence of a vehicle from water involves highly complex nonlinear fluid dynamics. The high velocity during this process causes complex cavitation phenomena. These cavitation bubbles grow and collapse as the vehicle moves, generating high-impact pressures that can threaten the vehicle's structural stability. Numerous factors influence the cavitation process. Among these, the shape of the vehicle's head is a particularly important factor, as it significantly influences the distribution of the surrounding flow field. Current research in this area relies primarily on experimental and numerical simulations using commercial software. Both of these are susceptible to environmental interference or waste of computing resources. To address this problem, this paper proposes a new method based on an improved U-Net neural network that, while only requiring input from coarse mesh results, can output results that would typically be obtained with a fine mesh. This method improves the traditional U-Net by introducing fast Fourier convolution and attention mechanisms, thereby avoiding the shortcomings of the U-Net such as the introduction of image noise in skip connections, the limitation of the convolutional perception area by the convolution kernel, and the inability to perceive globally. This reduces the root mean square error and the mean absolute error by about 40% even when flow field information is partially missing or there is noise, and improves the accuracy of flow field reconstruction by 5%–10%.
Zhao et al. (Wed,) studied this question.