Abstract: Transonic aerodynamic design is strongly influenced by nonlinear flow dynamics and complex engineering constraints, which makes the design process challenging and heavily reliant on designers’ experience and domain knowledge. The existing data-driven modeling and design methods have several challenges, i.e., poor generalization capability, insufficient interpretability, and limited ability to incorporate aerodynamic knowledge. To address these challenges, this paper proposes a flow-feature-oriented aerodynamic design agent, with a large language model serving as the core decision-making unit, to enable knowledge-driven and interpretable design of supercritical airfoils. In the proposed approach, airfoil geometric parameters and key flow features, such as wall Mach number distributions, are jointly represented as the state. Aerodynamic design knowledge and design history are introduced through structured prompts, guiding the model to generate geometry modification strategies that satisfy engineering constraints. Through interaction with a fast aerodynamic analysis environment, the model progressively establishes correlations among flow features, aerodynamic performance, and geometric modification actions, thereby realizing physics-guided aerodynamic design. A drag-reduction task for supercritical airfoils is used as the test case. The results demonstrate that the proposed agent achieves design performance comparable to greedy search methods, while exhibiting superior interpretability and stability in design process. Further comparisons between Chinese and English prompt, as well as large language models with different parameter scales, indicate that the task-specific fine-tuning process plays a significantly more important role in improving the performance of the aerodynamic design agent than the choice of prompt language itself. Overall, the proposed framework provides a new perspective for the development of trustworthy aerodynamic design agents for engineering applications.

Key words: transonic, aerodynamic design, supercritical airfoil, vision-language model, large language model, agent, reinforcement learning, flow feature