Correlating Structure with Dynamics in Supercooled Liquids Using Machine Learning Tools Restricted; Files & ToC

Abstract

Understanding the relationship between structure and dynamics in supercooled liquids remains a central challenge in glass physics. Machine learning techniques, particularly Support Vector Classification (SVC), have provided insights into structural predictors of dynamics, such as the “softness” order parameter, which correlates with energy barriers for particle rearrangements. This dissertation critically examines the methodology and interpretability of machine learning models in this context, focusing on their ability to predict rearrangement probabilities and energy barriers. We first investigate whether classification hyperplanes trained on structural data from high-temperature, diffusive regimes can predict energy barriers in the supercooled regime. By introducing a Z-score-based binning approach, we demonstrate that structural features associated with purely diffusive motion retain predictive power for activated events, challenging conventional assumptions about structure-dynamics correlations at high temperatures. Building on this, we explore the physical interpretability of various regression-based machine learning models, including Ridge Regression, Support Vector Regression, and Multilayer Perceptron, in predicting energy barriers. Our analysis, leveraging the iso-configurational ensembles, shows that these models capture similar structural signatures as SVC, reinforcing the idea that predictive success is rooted in the ability to learn high-dimensional structure-dynamics relationships. We then extend our investigation to the role of memory effects in supercooled liquids, using softness as a structural order parameter to probe system responses under thermal cycling. Preliminary findings suggest that supercooled liquids exhibit memory effects typically associated with glasses, raising new questions about the glass transition. This dissertation highlights key challenges in using machine learning to understand glassy dynamics, particularly regarding feature selection, model interpretability, and the physical significance of learned representations. Our findings contribute to ongoing efforts to develop more robust and interpretable machine learning frameworks for studying supercooled liquids and other complex systems. 

Table of Contents

This table of contents is under embargo until 22 November 2025

About this Dissertation

Rights statement

• Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.

School	Laney Graduate School
Department	Physics
Degree	Ph.D.
Submission	Dissertation
Language	English
Research Field	Psychology, General Physics, Molecular Physics, Condensed Matter
Stichwort	Supercooled liquids Machine Learning Glass physics
Committee Chair / Thesis Advisor	Daniel Sussman, Emory University
Committee Members	Carl Yang, Emory University Eric Weeks, Emory University Gordon Berman, Emory University Justin Burton, Emory University

Zuletzt geändert

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	File download under embargo until 22 November 2025	2025-04-15 08:51:23 -0400	File download under embargo until 22 November 2025

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions