Residual network intelligent prediction method for hypersonic inlet internal contraction basic flowfield

This study aims to establish a rapid prediction method for hypersonic inlet internal contraction basic flowfield to improve the design efficiency of inward-turning inlets. The focus of this study is to develop a reliable prediction model capable of quickly predicting flow field distributions of hypersonic inlet basic flowfield without conducting time-consuming CFD (computational fluid dynamics) simulations.

A parametric design approach using quasi-uniform B-spline methods was implemented to characterize the internal contraction basic flowfield. This parametric representation captured key geometric features, including initial compression angle, tangent angle at the end of outer compression, total contraction ratio, internal contraction ratio, and exit direction angle. Based on this parameterization, a comprehensive dataset of flowfield samples was generated through high-fidelity CFD simulations. A deep learning framework based on the residual neural network (ResNet) architecture was then developed to predict flowfield characteristics from geometric parameters. The network was structured with multiple residual blocks to effectively learn the complex relationships between inlet geometry parameters and flowfield properties. Image quality assessment metrics including PSNR (peak signal-to-noise ratio) and SSIM (structural similarity index ) were employed to quantitatively evaluate the prediction accuracy by comparing the neural network outputs with reference CFD results.

The established prediction model for hypersonic inlet internal contraction basic flowfield demonstrates high accuracy, with excellent reconstruction precision for both the overall flowfield and critical regions. From a quantitative assessment perspective, the test set samples show an overall average peak signal-to-noise ratio of 42.51 dB and an average structural similarity index of 0.9973, indicating high fidelity of the flowfield prediction model. For predictions focusing on critical flow regions, although the peak signal-to-noise ratio and structural similarity index decrease slightly, they still maintain high levels, demonstrating that the prediction model possesses strong predictive capability for critical regions. Furthermore, wall characteristics and shock wave shape distributions were extracted from selected test set samples and compared with actual conditions, showing a high degree of agreement. This confirms that the model also achieves high precision in predicting key flow field features.

Parametric design of internal contraction and centerbody basic flowfield was achieved based on quasi-uniform B-splines. This method enables parametric expression of key geometric features of the basic flowfield, such as initial compression angle, downwash angle, total contraction ratio, internal contraction ratio, and exit direction angle, providing an effective tool for basic flowfield sample generation. A rapid prediction model for hypersonic internal contraction basic flowfield was constructed based on a data-driven residual neural network architecture. This model demonstrates high accuracy in predicting flowfield characteristics and performance distributions, with an overall flowfield average peak signal-to-noise ratio of 42.51 dB and average structural similarity index above 0.99. It exhibits good fidelity for flowfield prediction within the design geometric parameter sample space. Image evaluation methods were used to assess key flow field features in the internal contraction basic flowfield. Results indicate that the prediction model can accurately capture shock wave shapes, wall characteristics distributions, with overall trends conforming to expectations, demonstrating good feature extraction and flowfield prediction capabilities.