Control of quasi-equilibrium state of annular flow through reinforcement learning

Stability control of the convection flow field has always been a focal issue. The annular discussed in this work is typical research model microgravity fluid physics, which extracted from industrial crystal growth by Czochralski method. It believed that instability thermal key factor affecting quality growth. Combining reinforcement learning algorithm with neural network, paper proposes policy makes forced compete thermocapillary changing dynamic boundary conditions system. This successfully applied to quasi-equilibrium state flow, and global stability well maintained. first experimentally under low medium Ma numbers achieve state, different before onset oscillations. Then, simulation environment created imitate experimental conditions. After training environment, self-optimized algorithm, machine approach can maintain for long period time. Finally, method validated completely optimized using same optimization similar network structure. demonstrates understand physical author's objectives through learning. an important application real world clear demonstration value physics.