Agentic AI for Computational Fluid Dynamics: A Review

Computational fluid dynamics (CFD) workflows remain strongly dependent on expert decisions in geometry preparation, meshing, model and solver configuration, convergence assessment, post-processing, and design iteration. This review examines agentic artificial intelligence (AI) for CFD automation and engineering design, with emphasis on reinforcement learning (RL) agents and large language model (LLM) agents. RL agents provide rapid, task-specific, and physics-aware execution within predefined observation and action spaces, whereas LLM agents offer flexible goal interpretation, knowledge access, planning, and tool orchestration. To organize these potentially complementary capabilities, a unified taxonomy is introduced based on decision paradigm, CFD role, and degree of autonomy, while multi-agent organization is treated as a separate architectural property. The reviewed evidence indicates that neither RL nor LLM agents alone are sufficient for trustworthy autonomous CFD. A hierarchical autonomous-CFD-engineer architecture is therefore proposed, combining an LLM orchestrator, domain-specialized agents, conventional CFD tools, independent numerical and physical verification, shared provenance, and human governance. Extending this architecture toward hypothesis formulation, numerical-experiment design, and model discovery suggests a pathway to an AI CFD scientist; however, current systems support bounded automation and supervised scientific collaboration rather than independent scientific judgment.