Machine learning glass caging order parameters with an artificial nested neural network.

Area of Science:

• Condensed Matter Physics
• Computational Materials Science
• Statistical Mechanics

Background:

• Supercooled liquids exhibit dramatic dynamic slowing and particle caging near the glass transition, while structural changes are minimal.
• Existing machine learning approaches often focus on structural predictors for glassy dynamics, overlooking direct dynamic feature learning.
• Understanding the nature of different glass transitions (melting, Gardner, liquid-to-glass) is crucial for materials science.

Purpose of the Study:

• To develop a machine learning method that learns particle caging features directly from dynamics, rather than relying on structural properties.
• To apply this method to identify and characterize order parameters for three distinct transitions in a hard sphere glass model.
• To classify the nature of these transitions (e.g., first-order, second-order, crossover) using a finite-size scaling analysis.

Main Methods:

• Implementation of a two-level nested neural network machine learning algorithm.
• Definition of particle caging features based purely on dynamic behavior.
• Application to a simulated hard sphere glass model, analyzing melting, Gardner, and liquid-to-ordinary glass transitions.

Main Results:

• The machine learning model successfully identified appropriate caging order parameters for all three investigated transitions.
• The algorithm achieved accurate phase classification for input samples across different glassy states.
• Finite-size scaling analysis of classification results revealed the melting transition as first-order and the Gardner transition as second-order.
• The liquid-to-ordinary glass transition was correctly identified as a dynamic crossover, not a sharp phase transition.

Conclusions:

• This study demonstrates a novel, generic machine learning approach for learning dynamical features in glassy systems.
• The method requires minimal system-specific knowledge, offering a powerful tool for studying complex dynamic phenomena.
• The findings provide new insights into the classification and understanding of different types of glass transitions.