Efficient deep reinforcement learning strategies for active flow control based on physics-informed neural networks.

This article focuses on a novel active flow control method that integrates physics-informed neural networks (PINNs) with deep reinforcement learning (DRL), termed physics-informed deep reinforcement learning (PIDRL), to reduce reliance on intrusive flow probes while enhancing control performance. Using a two-dimensional cylinder flow at Reynolds number 100 as a case study, the approach strategically employs a minimal number of probes (reduced from 164 to 4) combined with PINN-based flow field predictions to provide enriched input data for the DRL agent. Compared to traditional DRL, PIDRL achieves up to a 32% reduction in drag coefficient and a 75% reduction in lift coefficient fluctuations, with improved training stability and a 6% enhancement in control effectiveness after stabilization. The study highlights the importance of probe placement and sampling density, demonstrating that PIDRL’s use of physics-informed predictions diminishes sensitivity to probe location and reduces the need for dense sensor arrays, thereby offering a more cost-effective and robust flow control strategy.