Fluid mixing optimization with reinforcement learning.

Area of Science:

• Fluid Dynamics
• Computational Science
• Artificial Intelligence

Background:

• Fluid mixing is vital for industrial processes.
• Global-in-time optimization is key for complex mixing challenges.
• Passive scalar field mixing requires efficient optimization strategies.

Purpose of the Study:

• To apply reinforcement learning (RL) for optimizing passive scalar fluid mixing.
• To investigate the characteristics of RL suitable for global-in-time optimization in fluid dynamics.
• To develop a novel transfer learning approach for fluid mixers.

Main Methods:

• Utilized reinforcement learning (RL) to train a mixer for a 2D fluid mixing problem governed by advection-diffusion equations.
• Analyzed the dynamics of stretching and folding around stagnation points in the optimized mixing process.
• Developed and tested a transfer learning method to adapt mixers across different Péclet numbers.

Main Results:

• The trained RL mixer achieved exponentially fast mixing without prior knowledge.
• Stretching and folding mechanisms around stagnation points were identified as crucial for optimal mixing.
• A physically reasonable transfer learning method was successfully demonstrated for varying Péclet numbers.

Conclusions:

• Reinforcement learning is a powerful tool for optimizing fluid mixing, particularly for passive scalar fields.
• The developed transfer learning approach enhances the applicability of RL-trained mixers to different industrial conditions.
• The findings suggest potential applications in various industrial mixing scenarios, including turbulent mixing.