Aerodynamic design plays a vital role in rocket design. To expedite the design process of rocket aerodynamic shapes and minimize tedious iterative optimization, this study proposes a generative rapid inverse design method for rocket aerodynamic shapes. This method employs image-based geometric characterization of rocket aerodynamic shapes and integrates three main components: a conditional diffusion model-based rocket aerodynamic shape generation model, a convolutional neural network (CNN)-based rocket aerodynamic performance rapid prediction model, and a rocket design parameter discrimination model. The aerodynamic shape generation model takes the axial force coefficient C_A and the normal force coefficient C_N under typical operating conditions as inputs to rapidly generate numerous imaged rocket aerodynamic shape schemes. Subsequently, the performance prediction model screens the designs that meet the specified targets. Finally, the design parameter discrimination model extracts parameters from the selected scheme for validation or converts it into a three-dimensional shape. The dataset used in this study contains 165000 sets of rocket aerodynamic shapes and their corresponding aerodynamic performance. The prediction model was tested on a typical axisymmetric rocket configuration, requiring only about 30 seconds to predict 100 imaged rocket aerodynamic shapes. The optimal design exhibits remarkable accuracy, with deviations of only 0.0827% and 0.7124% for the axial and normal force coefficients, respectively. These results demonstrate the feasibility of the proposed method for rocket aerodynamic shape design and offer new insights for related research. Additionally, this paper analyzes the impact of the cross-attention mechanism and finds that it does not optimize model accuracy; this exploratory finding provides a reference for subsequent model optimization.

As one of the fastest-growing directions in artificial intelligence, the large model technology has achieved remarkable success in realms such as natural language processing and computer vision and is vigorously expanding its influence in empowering scientific research. It has also become a powerful tool in aerodynamics, possessing significant potential to expedite aerodynamic experiments and computations and assist aerodynamic theory and knowledge discovery. This paper begins by presenting an overview of large models for language processing, computer vision, and scientific computing. Subsequently, the paper outlines the conceptual framework of large models for scientific computing in aerodynamics, reviewing the current research progress from various perspectives, including flow field prediction, turbulence modeling, aerodynamic performance prediction, and aerodynamic configuration design. Furthermore, key techniques of large models in aerodynamics are discussed in-depth from the perspectives of model architecture, feedback alignment, and the generation of big aerodynamic data. Lastly, developing directions of large models in aerodynamics are prospected, including the construction of a unified pre-trained foundational model, the integration of aerodynamic knowledge to support scientific discoveries, and the development of discipline-specific agents.