Machine learning strategies for path-planning microswimmers in turbulent flows

Area of Science:

• Fluid dynamics
• Artificial intelligence
• Microswimmer locomotion

Background:

• Turbulent fluid flows present significant challenges for microswimmer navigation.
• Traditional strategies often prove inefficient for microswimmers in complex flows.

Purpose of the Study:

• To develop and evaluate an adversarial-reinforcement learning scheme for microswimmer navigation in turbulent flows.
• To determine if this scheme can enable microswimmers to reach targets more efficiently than naive strategies.

Main Methods:

• Utilized pseudospectral direct numerical simulations of Navier-Stokes equations to model turbulent flows in 2D and 3D.
• Introduced passive microswimmers into simulated flows, employing an adversarial-reinforcement learning control scheme.
• Analyzed the impact of microswimmer bare velocity to fluid velocity ratio and response time to vorticity product on navigation.

Main Results:

• The adversarial-reinforcement learning scheme enabled microswimmers to discover nontrivial paths.
• Microswimmers using the developed scheme reached targets on average in less time compared to naive swimmers.
• Navigation efficiency improved as the learning scheme adapted to the turbulent flow conditions.

Conclusions:

• Adversarial-reinforcement learning offers a promising approach for optimizing microswimmer trajectories in turbulent environments.
• The developed scheme demonstrates superior performance over basic strategies for target-reaching tasks.
• Control parameters significantly influence the effectiveness of the learning-based navigation strategy.