Reinforcement learning optimisation for graded metamaterial design using a physical-based constraint on the state...

Area of Science:

• Materials Science
• Mechanical Engineering
• Energy Harvesting

Background:

• Graded metamaterials offer tunable properties for energy harvesting applications.
• Piezo-mechanical systems are crucial for converting mechanical vibrations into electrical energy.
• Optimizing metamaterial design under realistic conditions is essential for efficient energy harvesting.

Purpose of the Study:

• To maximize the energy harvesting capability of a graded metamaterial using reinforcement learning (RL).
• To investigate the performance of RL-based optimization under realistic microscale excitations, including magnetic loading and random forces.
• To compare the effectiveness of RL with genetic algorithms for optimizing metamaterial grading rules.

Main Methods:

• Modeling the piezo-mechanical system using equivalent lumped parameters derived from general impedance analysis.
• Implementing a reinforcement learning (RL) approach to optimize the metamaterial's energy harvesting capability.
• Constraining the RL state representation and action space using physical insights into wave propagation.

Main Results:

• The RL-based optimization successfully maximized energy harvesting under realistic conditions.
• Genetic algorithms were more effective in deterministic settings (magnetic loading).
• RL demonstrated superior performance in stochastic settings with random excitations.

Conclusions:

• Reinforcement learning provides a powerful and adaptable tool for optimizing energy harvesting in graded metamaterials, particularly under complex, stochastic excitations.
• The integration of physical understanding into the RL framework enhances its efficiency and effectiveness.
• RL-based optimization presents a promising alternative to traditional methods like genetic algorithms for advanced material design.