Generalization of three-dimensional flight vehicle shape representation using generative models

The increasingly complex and dynamic design requirements of future hypersonic vehicle have heightened the call for enhanced flexibility and efficacy in aerodynamic shape parameterization methods. This research addresses the constraints imposed by conventional geometric parameterization methods in hypersonic vehicle design, which tend to confine aerodynamic shapes excessively, limiting design optimization and culminating in the “dimensional disaster”. To tackle these challenges, the study introduces a novel approach centered on generative models to generalize the depiction of three-dimensional hypersonic vehicle aerodynamic profiles. This innovative method entails extracting geometric attributes from aerodynamic profile images and imposing constraints on global and local point cloud diffusion models to facilitate the creation of three-dimensional point cloud representations, catering to the intricate demands of three-dimensional shape reduction and adaptable design in engineering contexts. The method achieves a comprehensive multi-view generalized representation pipeline from cross-sectional images to point cloud geometry and surface meshes through three core modules: a geometric feature extraction and dimensionality reduction reconstruction model for cross-sectional images based on Variational Autoencoder (VAE) and residual networks; a 3D geometric point cloud generation model integrating point cloud variational autoencoder networks with conditionally controllable generative diffusion models; a surface mesh reconstruction model employing a differentiable Poisson solver. In the context of the three-dimensional blended wing-body configuration, manipulating generalized shape parameters within latent variable spaces enables rapid generation of single profile images in 3 s and three-dimensional aerodynamic mesh files in 80 s. The reconstructed shapes maintain an average geometric error within 1 mm compared to the initial design, demonstrating efficient and accurate generalization of complex aircraft shapes. Furthermore, the generated surface meshes were directly imported into Computational Fluid Dynamics (CFD) software to predict flow fields and aerodynamic characteristics under different angles of attack. The results highlight the potential of this method as an intelligent and flexible aerodynamic shape generalization tool for optimizing complex aircraft designs.