Predicting Dynamic Heterogeneity in Glass-Forming Liquids by Physics-Inspired Machine Learning

Area of Science:

• Condensed Matter Physics
• Materials Science
• Computational Chemistry

Background:

• Deeply supercooled liquids exhibit complex dynamics crucial for understanding glass transitions.
• Predicting these dynamics often requires extensive simulations and computationally expensive methods.
• Current machine learning approaches may lack accuracy or require substantial training data.

Purpose of the Study:

• To introduce GlassMLP, a novel machine learning framework for predicting long-time dynamics in deeply supercooled liquids.
• To demonstrate the framework's ability to utilize physics-inspired structural input for enhanced predictive power.
• To achieve state-of-the-art performance with improved data and parameter efficiency.

Main Methods:

• Development of a deep neural network architecture, GlassMLP.
• Incorporation of physics-inspired structural features as input.
• Application to atomistic models in both 2D and 3D systems.
• Quantitative prediction of four-point dynamic correlations and dynamic heterogeneity geometry.

Main Results:

• GlassMLP outperforms existing state-of-the-art methods in predicting liquid dynamics.
• The framework demonstrates superior parsimony in terms of training data and fitting parameters.
• Accurate quantitative prediction of four-point dynamic correlations and dynamic heterogeneity.
• Successful transferability across different system sizes, enabling efficient temperature-dependent analysis.

Conclusions:

• GlassMLP offers a powerful and efficient approach for modeling supercooled liquid dynamics.
• The use of physics-inspired inputs significantly enhances predictive accuracy and data efficiency.
• The study reveals critical temperature-dependent changes in the geometry of rearranging regions within these liquids.