Curiosity-driven search for novel nonequilibrium behaviors.

Area of Science:

• Complex Systems Science
• Machine Learning
• Statistical Physics

Background:

• Exploring the full range of behaviors in complex systems is challenging.
• Existing sampling techniques often require predefined order parameters, which are unknown for many nonequilibrium systems.
• This creates a 'chicken-and-egg' problem where new behaviors are needed to define parameters, but parameters are needed to find new behaviors.

Purpose of the Study:

• To develop an automated method for exploring nonequilibrium systems with unknown order parameters.
• To combine active learning and unsupervised learning for efficient behavior discovery.
• To address the challenge of identifying novel behaviors and their corresponding order parameters.

Main Methods:

• Implemented an iterative approach combining active learning and unsupervised learning.
• Active learning was used to expand the library of observed behaviors based on current order parameters.
• Unsupervised learning was employed to relearn and refine order parameters from the expanded behavior library.

Main Results:

• Successfully demonstrated the approach on Kuramoto models of increasing complexity.
• Reproduced known system phases and identified previously unknown behaviors.
• Discovered new order parameters associated with these novel behaviors.

Conclusions:

• The proposed active and unsupervised learning framework effectively automates the exploration of complex nonequilibrium systems.
• It overcomes the limitation of unknown order parameters, enabling the discovery of new system behaviors.
• The method provides a pathway to align automated search with human intuition in complex system exploration.