Research on intelligent prediction method of supersonic flow field in scramjet based on deep learning: A review

Direct numerical simulation (DNS) serves as a crucial method for optimizing and validating scramjets, significantly advancing their design process. Nonetheless, solving the Navier-Stokes equations numerically entails substantial computational expenses, particularly for large-scale projects characterized by intricate hysteresis and high-precision thermochemical reactions. In recent years, numerous studies have demonstrated the rationality and efficacy of deep learning in reconstructing the evolutionary characteristics of flow fields. To reduce neural network models' dependence on high-fidelity data and prevent the generation of non-physical solutions, neural networks incorporating physical information constraints offer a novel learning paradigm. This approach encodes prior knowledge and physical interpretability, which traditional neural networks lack. Based on this, this study investigates, analyzes, and summarizes traditional prediction methods for supersonic combustion flow, datadriven intelligent solution algorithms for supersonic flow fields, lightweight neural network models, and intelligent prediction algorithms for flow fields using physical information neural networks.