关键词: 高速飞行器, D-最优试验设计, 迁移融合, 数据融合, 预训练微调

Abstract: To address the high costs of obtaining high-fidelity data in aerodynamic design and the insufficient prediction accuracy of traditional surrogate models in strong nonlinear regions such as high angles of attack and transonic speeds, this paper proposes a multi-fidelity aerodynamic data fusion modeling framework coupling the D-optimal experimental design criterion with deep transfer learning. The method first constructs a low-fidelity dataset covering the entire flight envelope using low-cost Euler simulations to rapidly capture the global physical trends of aerodynamic forces. Based on this, the D-optimal criterion is introduced to actively select key operating points that contribute most to model parameter estimation by maximizing the determinant of the information matrix, followed by high-fidelity numerical simulations based on the Reynolds-Averaged Navier-Stokes equations.Regarding the model architecture, a deep neural network based on pre-training and fine-tuning mechanisms is established, utilizing a large volume of low-fidelity data to train feature extraction layers for building a general flow field feature space. A parameter freezing strategy is then employed, using a minimal number of high-fidelity samples to fine-tune only the output layer weights, thereby achieving precise mapping from low-fidelity trends to high-fidelity values. Taking a typical three-dimensional missile as the research object, validation was conducted within a broad envelope ranging from Mach 1.5 to 6 and angles of attack from minus 2 degrees to 20 degrees.The research results indicate that the model fine-tuned using the D-optimal sampling strategy reduces the root mean square error on the test set by approximately 6.41 percent compared to Latin Hypercube Sampling, demonstrating a superior ability to capture nonlinear features in strong shock interference zones. Compared to direct Kriging modeling, the fusion model reduces the mean square error for drag coefficient prediction by 78.1 percent and exhibits excellent fitting robustness across the entire range of operating conditions. Under the premise of maintaining equivalent prediction accuracy, this scheme saves approximately 13.33 percent of the total time cost compared to pure high-fidelity computation. This study proves the effectiveness of the proposed framework in handling high-dimensional and small-sample aerodynamic modeling problems, providing technical support for the efficient aerodynamic assessment and design optimization of complex aircraft.

Key words: High-speed Aerial Vehicle, D-optimal Design, Transfer Fusion, Data Fusion, Pre-training and Fine-tuning