Deep reinforcement cross-domain transfer learning of active flow control for three-dimensional bluff body flow

Abstract

This paper applies mutual information-based knowledge transfer learning with soft actor-critic (MIKT-SAC) algorithm to address cross-domain issues in state and action dimensions for active flow control (AFC). It explores the potential of deep reinforcement learning (DRL) in discovering novel drag reduction strategies. The algorithm starts with a pretrained agent on a two-dimensional (2D) case, extracting knowledge to mitigate aerodynamic forces acting on a 3D bluff body under high Reynolds number flow conditions. The algorithm is applied to two test cases to demonstrate its capabilities and limits: The first investigates the state dimension mismatch problem using a 3D square cylinder at high Reynolds number $Re=22000$, where four jets at the corners of square cylinder as actuators. The second test examines scenarios with both state and action dimension mismatches using a circular cylinder with multiple zero-net-mass-flux jets positioned as two slots on the top and bottom surfaces. The results show that MIKT-SAC method outperforms the vanilla SAC algorithm, significantly reducing 51.1% and 45.1% training time and reducing drag coefficients ($C_D$) by 50.9% and 49.4% for the square and circular cylinders, respectively, while effectively suppressing drag and lift fluctuations. The multi-jet actuation delays vortex shedding on the surface of bluff body, reducing fluctuating lift forces on both cases. These findings highlight the potential of DRL in active flow control, laying a foundation for efficient, robust, and practical implementation of bluff body control technologies in practical engineering applications.