A numerical study of fish adaption behaviors in complex environments with a deep reinforcement learning and immersed boundary-lattice Boltzmann method

Fish adaption behaviors in complex environments are of great importance in improving the performance of underwater vehicles. This work presents a numerical study of the adaption behaviors of self-propelled fish in complex environments by developing a numerical framework of deep learning and immersed boundary-lattice Boltzmann method (IB-LBM). In this framework, the fish swimming in a viscous incompressible flow is simulated with an IB-LBM which is validated by conducting two benchmark problems including a uniform flow over a stationary cylinder and a self-propelled anguilliform swimming in a quiescent flow. Furthermore, a deep recurrent Q-network (DRQN) is incorporated with the IB-LBM to train the fish model to adapt its motion to optimally achieve a specific task, such as prey capture, rheotaxis and Karman gaiting. Compared to existing learning models for fish, this work incorporates the fish position, velocity and acceleration into the state space in the DRQN; and it considers the amplitude and frequency action spaces as well as the historical effects. This framework makes use of the high computational efficiency of the IB-LBM which is of crucial importance for the effective coupling with learning algorithms. Applications of the proposed numerical framework in point-to-point swimming in quiescent flow and position holding both in a uniform stream and a Karman vortex street demonstrate the strategies used to adapt to different situations.

Automated summary

This study presents a numerical framework utilizing deep learning and IB-LBM to study fish adaptation behaviors in complex environments. The fish behavior is validated through benchmark problems and trained to adapt its motion for specific tasks. The framework efficiently couples with learning algorithms and demonstrates strategies for adaptation in different situations.