Predicting dynamic heterogeneity in glass-forming liquids by physics-informed machine learning - Physics-informed Machine Learning

Aug 4, 2026

Upload by Himbo

Introduces GlassMLP, a physics-informed deep learning framework designed to predict long-time dynamics in deeply supercooled glass-forming liquids. Using established structural order parameters as input, the method is applied to atomistic models in both 2D and 3D, achieving improved performance over prior approaches with fewer training data and fitting parameters. It quantitatively predicts four-point dynamic correlations and the geometry of dynamic heterogeneity, and its transferability enables tracking temperature evolution of spatial correlations and rearranging-region geometry.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages12

Reading Time30 min

AccessFree to download

GlassMLP framework

Physics-informed inputs and prediction targets

Model systems and simulation setup

Propensity-based dynamics analysis

Temperature evolution of dynamic heterogeneity

arXiv :2210 . 16623v2 [ cond-mat .soft] 4 Jan 2023 Predicting dynamic heterogeneity in glass-forming liquids by physics-informed machine learning Gerhard Jung, 1 Giulio Biroli,2 and Ludovic Berthier 1, 3 1 Laboratoire Charles Coulomb (L2C), Universit􀀓e de Montpellier, CNRS, 34095 Montpellier, France 2 Laboratoire de Physique de l'Ecole Normale Sup􀀓erieure, ENS, Universit􀀓e PSL, CNRS, Sorbonne Universit􀀓e, Universit􀀓e de Paris, F-75005 Paris, France 3 Yusuf Hamied Department of Chemistry, University of Cambridge, Lens􀀌eld Road, Cambridge CB2 1EW, United Kingdom (Dated: January 5, 2023) We introduce GlassMLP, a machine learning framework using physics-informed structural input to predict the long-time dynamics in deeply supercooled liquids. We apply this deep neural network to atomistic models in 2D and 3D. Its performance is better than the state of the art while being more parsimonious in terms of training data and 􀀌tting parameters. GlassMLP quantitatively predicts four-point dynamic correlations and the geometry of dynamic heterogeneity. Its transferability from small to large system sizes allows us to probe the temperature evolution of spatial dynamic correlations, revealing a profound change with temperature in the geometry of rearranging regions. Glasses are formed by the continuous solidi􀀌cation of supercooled liquids under cooling, while maintaining an amorphous microstructure [1] . They are fascinating as they combine the complex properties of solids and liquids [2] . Understanding glass formation and the phenomenon of the glass transition has been the focus of an intense research activity [3] . An important characteristic of supercooled liquids is the emergence and growth of spatial heterogeneity characterising the relaxation dynamics, where some regions actively rearrange while other appear completely frozen [4] . In recent years, an important e􀀋ort was devoted to understanding the connection between dynamic heterogeneity and structural properties [5, 6] . Several structural order parameters were shown to correlate with the dynamics, including density, potential energy [7], locally favored structures [8{10], but also more complicated quantities such as soft modes [11], local yield stress [12] and Franz-Parisi potential [13] . The search intensi􀀌ed with the emergence of machine learning (ML) allowing the detection of correlations from unsupervised [14{16] or supervised [17{23] learning. The explored methodologies range from simple linear regression and support vector machines using a set of handcrafted structural descriptors [17] to graph neural networks (GNN) with tens of thousands of adjustable parameters [19, 23] . Despite this versatility, none of the proposed networks can currently predict dynamic heterogeneities and related multipoint correlation functions that quantitatively agree with the actual dynamics. This is an open challenge because predictability of the dynamics from the structure is weak at the single particle level and only becomes meaningful at larger length scales [24] . Here, we bridge this major gap by leveraging previous ML approaches and combining them. We introduce a physics-informed deep neural network that uses established structural order parameters as input to predict long-time dynamics in deeply supercooled liquids. The proposed methodology, which surpasses the state of theart, allows us to obtain quantitative predictions about spatially heterogeneous dynamics and hence to gather physical insights about their temperature evolution. We simulate a Lennard-Jones non-additive mixture in 3D (KA,[25]) for comparison with earlier work [19] anda 2D ternary mixture (KA2D) . We focus on KA2D since its interactions were adapted to e􀀎ciently prevent crystallization [26] and enable the use of the swap Monte Carlo (SWAP) algorithm [27, 28], allowing us to analyse very low temperatures. Equilibrium con􀀌gurations are created with N = 1290 particles (Mtype = 3, N1 = 600, N2 =

FAQs

What is GlassMLP and what problem does it address?

GlassMLP is a physics-informed machine learning framework that predicts long-time dynamics in deeply supercooled glass-forming liquids, focusing on dynamic heterogeneity and multipoint correlations.

How does the method connect structure to dynamics?

It uses established structural order parameters as physics-informed inputs and trains a deep neural network to predict dynamics characterized by propensity and related relaxation behavior.

What quantities does the framework predict quantitatively?

It quantitatively predicts four-point dynamic correlations and the geometry of dynamic heterogeneity, allowing analysis of how rearranging regions change with temperature.

Aug 4, 2026

Upload by Himbo