Computational fluid dynamics (CFD) is essential for advancing scientific and engineering fields, but is hindered by operational complexity, high expertise requirements, and limited accessibility. This article introduces ChatCFD, a large language model (LLM) ‐driven agent system for end‐to‐end CFD automation. Powered by DeepSeek‐R1/V3, a multiagent architecture, structured OpenFOAM knowledge bases, precise error locator, and iterative reflection, ChatCFD dramatically outperforms prior systems. On 315 benchmark cases, it attains 82. 1% execution success (vs. 6. 2% MetaOpenFOAM, 42. 3% Foam‐Agent) and, crucially, 68. 12% physical fidelity—the first rigorous metric capturing whether a runnable simulation is scientifically meaningful. A dedicated Physics Interpreter achieves 97. 4% summary fidelity, exposing the striking LLM gap between fluent narration and enforcing dozens of tightly coupled physical constraints in executable code. Resource‐efficiency analysis highlights ChatCFD's practical advantage, achieving an average of 192. 1 k tokens and 0. 208 per case—half the tokens of Foam‐Agent and 1. 5 cheaper than MetaOpenFOAM. Ablation studies confirm that structured domain knowledge and reasoning is indispensable: removing the solver template DB collapses accuracy to 48%, while the error locator module proves the single most critical component. Flexibility experiments show autonomous solver selection across compressible/incompressible and steady/transient regimes (95. 23% success) and turbulence‐model switching (100% success), even on unseen configurations. By faithfully reproducing complex literature cases NACA0012 airfoil and supersonic nozzle with 60%–80% end‐to‐end success where baselines fail entirely, ChatCFD establishes new benchmarks for artificial intelligence (AI) ‐driven CFD automation. Physics coupling analyses reveal higher resource demands for multiphysics‐coupled cases, while LLM bias toward simpler setups introduces persistent errors. ChatCFD's modular, MCP‐compatible design directly enables collaborative multiagent networks and paves the way for scalable AI‐driven CFD innovation. The code for ChatCFD is available at: https: //github. com/ConMoo/ChatCFD.
Fan et al. (Mon,) studied this question.