Supervised learning through physical changes in a mechanical system.

Area of Science:

• Materials Science
• Mechanical Engineering
• Robotics
• Machine Learning

Background:

• Mechanical metamaterials typically require precise force-response specifications.
• Real-world applications often lack exact specifications but provide example data.
• Learning frameworks are needed for adaptable mechanical systems.

Purpose of the Study:

• To develop a supervised learning framework for thin, creased sheets.
• To enable these sheets to learn desired force-response behaviors through physical experience.
• To achieve generalization to previously unseen forces.

Main Methods:

• A framework for supervised learning in thin, creased sheets was proposed.
• Sheets were trained by folding with example forces, altering local crease stiffnesses.
• The relationship between training error, test error, and sheet size was analyzed.

Main Results:

• The learning process reshaped inherent nonlinearities in the sheets.
• Sheets demonstrated correct responses to previously unseen test forces.
• The study established parallels between physical learning sheets and machine learning algorithms.

Conclusions:

• Physical learning in mechanical metamaterials can sculpt energy landscapes for desired behaviors.
• This framework offers a method for designing adaptable mechanical systems.
• Local physical learning processes can achieve complex, desired force-response relationships.