A novel framework for cost-effectively reconstructing the global flow field by super-resolution
Longyan Wang, Zhaohui Luo, Jian Xu, Wei Luo, Jianping Yuan
Abstract
Fluid data are of great significance for analyzing the fluid structure and understanding the law of fluid movement. Apart from the experimental test, the computational fluid dynamics (CFD) method has been widely applied in the field of fluid dynamics over the past few decades. However, due to the high computational costs of CFD method and the limitation of computational resources, it is still challenging to accurately calculate and obtain the high-resolution (HR) flow fields. To this end, a novel framework based on the super-resolution (SR) algorithm, namely, new enhanced down-sampled skip-connection and multi-scale (E-DSC/MS), is reported to achieve the HR global flow reconstruction from low-resolution data. Through the new SR flow reconstruction method, the HR flow fields of two benchmark 2D cases (i.e., cylinder and hydrofoil) are precisely and efficiently predicted using a universal SR model. The effectiveness of the new E-DSC/MS algorithm is tested by comparing it with the traditional super-resolution convolution neural network and U-net in terms of the velocity field prediction of the self-region (training region) and other-region (untrained region). The result shows that the universal SR flow reconstruction framework is able to increase the spatial resolution of velocity field by 16 times, and flow fields reconstructed by E-DSC/MS are in good agreement with the ground-truth data. In addition, the E-DSC/MS model could reconstruct the global flow field with a correlation coefficient of more than 99% regardless of the selection of the arbitrary region/window for SR training. The present method overcomes the limitation of the existing techniques in efficiently reconstructing HR flow field, which helps to reduce the requirement for expensive experimental equipment and to accelerate the CFD simulation process.