Physics-informed model-based deep reinforcement learning for dynamic flow control*

Model-based reinforcement learning (MBRL) is believed to have much higher sample efficiency compared to model-free algorithms by learning a predictive model of the environment. However, the performance of MBRL highly relies on the quality of the learned model, which is usually built in a black-box manner and may have poor predictive accuracy outside of the data distribution. The deficiencies of the learned model may prevent the policy from being fully optimized. In this work, we developed a physics-informed MBRL framework, where governing equations and physical constraints are utilized to inform the model learning and policy search. By incorporating the prior information of the environment, the quality of the learned model can be notably improved, while the required interactions with the environment are significantly reduced, leading to better sample efficiency and learning performance. Moreover, to effectively capture the long-span transition dynamics of fluid flow with irregular domains, a novel network architecture based on graph embedding and attention-based transformer is developed, and the effectiveness has been demonstrated on both incompressible and compressible flows.