Progressive Intelligent Framework for Supersonic Flow Separation Region Identification in Scramjets

This paper introduces a progressive intelligent framework to address key challenges in identifying flow separation regions in supersonic flow, including blurred boundaries, complex dynamic evolution, and limited identification accuracy. The framework integrates flow feature prediction and region segmentation into a unified process. First, a multipath fusion generative adversarial network is designed to establish a deep mapping relation between discrete wall-pressure measurements and the global flowfield structure, enabling accurate prediction of transient flowfields. Then, an end-to-end pixel-level segmentation model is applied to precisely extract separation region features from the predicted flow data. This progressive approach allows accurate identification of complex flow structures and separation zones under given constraints. Experimental results show that the predicted flowfields achieve peak signal-to-noise ratio, structure similarity index measure, and correlation coefficient values of 36.773, 0.985, and 0.991, respectively, whereas the identified separation regions reach an intersection over union of 0.956, a Dice coefficient of 0.975, and a recall of 0.980. All metrics significantly outperform those of comparative methods, demonstrating the framework’s practical utility for rapid diagnosis of complex flow structures.