A survey of large-language-model agents in aerodynamics

Large language model agents are driving a paradigm shift in aerodynamics from experience-driven approaches toward the triple integration of data, knowledge, and physics. To address the challenges of tacit knowledge inheritance and trusted collaborative execution of complex tasks, these agents establish an enhanced intelligence closed loop through two pathways: internalization of disciplinary knowledge, which enables structured sedimentation and reuse of unstructured knowledge, and intelligent orchestration of toolchains, which ensures physical consistency and procedural reliability in multi-tool collaboration. This review systematically surveyed the research progress of large language model agents in aerodynamics, constructed a five-dimensional technical framework consisting of prompt engineering and chain-of-thought reasoning, knowledge and tool augmentation, memory enhancement and long-context mechanisms, multi-agent systems, and human-in-the-loop. It reveals a three-stage evolutionary logic from automation through refinement to deep integration, identifies key technical gaps including the lack of evaluation systems, incomplete toolchain coverage, weak interdisciplinary knowledge fusion, insufficient human-machine collaboration depth, and computational efficiency optimization, and proposes development principles centered on discipline logic-led, technology capability-supported, and engineering value closed-loop. The review further envisions future directions such as collaborative evaluation ecosystem construction, deep internalization of physical laws, and efficient reasoning mechanisms. This review aims to provide systematic theoretical anchors and practical pathways for research on aerodynamic agents, foster interdisciplinary dialogue between artificial intelligence and aerodynamics, support equipment innovation and scientific breakthroughs, and advance the construction of a new intelligent aerodynamics paradigm.