How Large Language Models Can Enable Data-Driven Flow Control?

Although large language models (LLMs) have achieved remarkable success across a wide range of domains, their application in fluid mechanics remains relatively unexplored. Most existing research focuses primarily on language understanding and text generation. In this work, we investigate flow control of a rotating cylinder using reinforcement learning guided by language-based rewards. Leveraging WaterLily, an efficient numerical simulation library, we built a coupled RL-Fluid-LLM framework that enables the agent to explore based on intrinsic rewards derived from language. Our approach successfully reduced the drag coefficient to near zero. Results demonstrate that LLMs can effectively serve as semantic reward functions, enabling their integration into control frameworks for fluid dynamic systems.