Deep reinforcement learning control unlocks enhanced heat transfer in turbulent convection

Turbulent convection governs heat transport in both natural and industrial settings, yet optimizing it under extreme conditions remains a significant challenge. Traditional control strategies, such as predefined temperature modulation, struggle to achieve substantial enhancement. Here, we introduce a deep reinforcement learning (DRL) framework that autonomously discovers optimal control policies to maximize heat transfer in turbulent Rayleigh-Bénard convection. By dynamically adjusting wall temperature fluctuations, the DRL agent achieves a heat transfer enhancement of up to 38.5%, exceeding the 20 to 25% limit of conventional methods. The learned strategy reveals a nonlinear state-action relationship, inducing a fully modulated boundary layer regime. Furthermore, we distill the DRL insights into a simplified bang-bang control model, which retains comparable performance (up to 40.0% enhancement) and, crucially, generalizes to unseen, higher Rayleigh number cases without additional training. Our results demonstrate the power of machine learning in turbulence control and reveal a framework with potential for intelligent heat transfer optimization in real-world applications.