Finding efficient swimming strategies in a three-dimensional chaotic flow by reinforcement learning

Area of Science:

• Fluid dynamics
• Artificial intelligence
• Complex systems

Background:

• Microswimmers navigate complex three-dimensional fluid flows.
• Chaotic advection in Arnold-Beltrami-Childress flows presents navigation challenges.
• Optimal pathfinding in such environments is computationally demanding.

Purpose of the Study:

• To investigate the application of reinforcement learning for microswimmer navigation.
• To demonstrate that limited action sets can yield effective navigation strategies.
• To explore the efficiency of learning-from-experience in complex fluid environments.

Main Methods:

• A reinforcement learning algorithm was applied to point-like particles.
• Particles were simulated in a stationary superposition of two Arnold-Beltrami-Childress flows.
• Particles could choose from six fixed swimming directions with constant velocity amplitude.

Main Results:

• Particles learned approximately optimal strategies to navigate complex flows.
• The algorithm enabled swimmers to escape local fluid traps and move upward.
• A restricted set of actions proved sufficient for efficient strategy discovery.

Conclusions:

• Reinforcement learning is effective for learning complex navigation tasks in fluids.
• AI-driven agents can develop efficient strategies in chaotic environments with limited control.
• This approach highlights the potential of AI in fluid dynamics and robotics.