MAPping dynamic heterogeneity in supercooled glass-formers

Area of Science:

• Physics
• Materials Science
• Computational Science

Background:

• Dynamic heterogeneity in supercooled liquids is a key area of study in glass physics.
• Existing machine-learning models for predicting liquid dynamics often require labeled training data or lack interpretability.

Purpose of the Study:

• To develop an unsupervised machine-learning approach to infer dynamic heterogeneity from static structures in supercooled liquids.
• To introduce an intuitive method for characterizing dynamic heterogeneity and rare events in glass-formers.

Main Methods:

• Utilized an autoregressive generative model, Memory-Augmented Prediction (MAP), to learn particle configuration probabilities conditioned on local environments.
• Employed molecular-dynamics data for the supercooled Kob-Andersen binary system.
• Analyzed the MAP-derived reaction coordinate (Ω) to identify dynamic excitations and rare events.

Main Results:

• MAP successfully reproduced key benchmarks, including pair-correlation functions and point-to-set distance scaling.
• Modulations in the MAP-derived reaction coordinate (Ω) correlated with dynamic excitations.
• Tuning Ω revealed its sensitivity to local dynamic events, characterizing them as concurrent nearby excitations.

Conclusions:

• The unsupervised MAP model provides a novel and intuitive method to diagnose dynamic heterogeneity in supercooled liquids.
• MAP links rare dynamic events to atypical local structural states, offering new avenues for exploring glass-former dynamics.