基于深度强化学习与流固耦合技术的鱼类自主游动行为模拟.

The simulation of fish autonomous swimming has always been an important problem concerned by many disciplines, such as bionics, fish behavior and ecological hydraulics. Based on fluid structure coupling numerical simulation technology and deep reinforcement learning algorithm, an intelligent fish adaption behavior decision-making platform was established, which can realize the fish adaption swimming with the optimal decision-making scheme under different surrounding environment conditions. Deep reinforcement learning method was used to realize fish brain function, which simulating its continuous learning and final decision-making. The flow field and fish motion were simulated by immersed boundary-Lattice Boltzmann method to provide rich training samples for fish and execute fish brain decision. Based on this platform, the typical prey capture and Karman gait of fish were simulated and the training effect were analyzed. The simulation results show that in the prey capture simulating, the fish with different initial position angles can reach the target point with the optimal trajectory. In Karman gaiting, the fish tail beat frequency can be adjusted to approach the vortex shedding frequency, so as to absorb energy from the Karman vortex field to stabilize the gait in the vortex street. In the research of fish adaption swimming, the decision-making platform has stronger adaptability to complex flow field than traditional physical experiments, and can provide technical support for digital twinning in the fields of hydraulic engineering, ecological environmental engineering. [ABSTRACT FROM AUTHOR]