Active control strategies for spatiotemporally varying gas injection in a fluidized bed system are developed to optimize particle mixing using deep reinforcement learning (DRL). Unlike conventional pulsation methods relying on predefined waveforms with manually tuned parameters, the proposed framework autonomously discovers control policies through interaction between a DRL agent and a computational fluid dynamics (CFD) environment. The agent receives local voidage as the state and modulates gas velocities at three inlet segments as actions. A reward function is designed to simultaneously promote mixing uniformity, reduce power consumption, and maintain the fluidization regime. To alleviate the computational burden of CFD-based learning, transfer learning across grid resolutions and parallelized simulation environments is employed. Without any prior encoding of pulsation characteristics, the agent successfully discovers an in-phase sinusoidal injection strategy and further identifies a non-intuitive policy involving a reduced centerline velocity, which is unlikely to emerge from conventional predefined waveform-based approaches. The effectiveness of the learned policies is validated, and the underlying physical mechanisms are systematically analyzed. Whereas conventional pulsation alone improves mixing uniformity without noticeable energy savings, the DRL strategy improves mixing performance by an additional 42.9%, while simultaneously reducing power consumption by 3.37%. Such additional gains are attributed to reduced total gas input and the introduction of spatial asymmetry, compensating for the velocity deficit near the sidewalls due to the no-slip condition, enhancing lateral mixing.
Kim et al. (2025) studied this question.