Data-driven super-resolution reconstruction of supersonic flow field by convolutional neural networks

Abstract

The pursuit of high-resolution flow fields is meaningful for the development of hypersonic technology. Flow field super-resolution (SR) based on deep learning is a novel and effective method to provide HR flow fields in a scramjet isolator. Single-path and multiple-path network models based on convolutional neural networks (CNNs) have been developed to augment the spatial resolution of the experimental supersonic flow field. The single-path model uses a simple convolutional layer and fully connected layer serial architecture, and the multiple-path model increases the branch path by adding pooling layers to achieve a fusion structure architecture. Ground experiments of flow in a supersonic isolator at various working conditions are conducted to establish an experimental dataset. The trained single-path and multiple-path CNNs are compared with the traditional interpolation method on the flow field SR reconstruction accuracy. The results demonstrated that single-path CNNs have certain learning ability, but the SR accuracy is not satisfactory; multiple-path CNNs significantly improve the accuracy of flow field SR, and the multiple-path CNN with one branch path achieves the best SR performance.