Model-based reinforcement learning for active flow control: escalating performance and accelerated training

Active flow control (AFC) technology is crucial for achieving predominant drag reduction in marine engineering. To tackle the issue of substantial time consumption inherent in reinforcement learning (RL)-based AFC, this study presents a model-based reinforcement learning (MBRL) approach for active flow control around a circular cylinder. The surrogate models covering Re = 100–1000 under different conditions are constructed to enhance the control strategy and reduce training time, replacing the traditional computational fluid dynamics (CFD)-based RL environment. The MBRL-based strategy achieved drag reduction rates of 14.6%, 25.4%, 31.2%, and 21.8% at Re of 100, 200, 500, and 1000 respectively, demonstrating the effectiveness of the model-based reinforcement learning method in active flow control. Considering the data scarcity and complexity in predicting high Reynolds number flow fields, a transfer training strategy is implemented. Upon retraining and fine-tuning at Re = 1000, the drag reduction increased to 36.2%. Meanwhile, the proposed method achieves significantly faster policy optimisation compared to single-thread implementations, reducing the time required for convergence by orders of magnitude. This study advances the application of deep reinforcement learning in fluid dynamics by establishing a model-based framework. The MBRL method effectively adapts to complex flow field changes, facilitates flow performance optimisation, and holds promise for future implementation of active flow control technology in diverse and intricate fluid flow environments.