Geometric encoding of thermal history in glasses: Strain topology as a learnable structural signature
Year: 2026
Type: article
Abstract: How does cooling rate become encoded in the atomic structure of a glass? We demonstrate that thermal history is geometrically encoded in the strain topology —the spatial distribution of local mechanical distortions—rather than in scalar density or coordination metrics. Using GATv2 graph neural networks as structural probes with systematic ablation on a Lennard-Jones model system, we achieve classification accuracy between fast- and slow-cooled glasses under 5-fold stratified cross-validation. Pure topological features fail at near-random accuracy ( ), while bond-length statistics and local Hessian eigenvalues together carry the discriminative signal; the signal survives per-sample instance normalisation ( ), confirming that memory is relational rather than a global stiffness offset. To assess generality, we apply the identical GATv2 framework to the Kob-Andersen binary mixture using GlassBench configurations ( ). Using 5D pure-geometry bond statistics with no Hessian input and config-level stratified 5-fold cross-validation (all patches from a single parent configuration held within the same fold), we achieve accuracy (AUC , F1 ). An ablation removing the Hessian entirely yields , confirming that pairwise bond-length statistics are a sufficient geometric descriptor for thermal-history classification in the binary mixture. Beyond classification, we train GATv2 as a regressor across eight logarithmically spaced cooling rates spanning 3.5 decades. The first principal component of the 64-dimensional latent space captures of all structural variance and is monotonically ordered with cooling rate across six of eight rates (Spearman , ), constituting a structural proxy for fictive temperature. The geometric signal is more sensitive to cooling rate than potential energy per particle. Finally, geometric memory collapses abruptly within the first 50 Brownian scrambling steps at , consistent with rapid structural decorrelation at high temperature, and slow-cooled glasses develop extended percolating force chains absent in fast-cooled counterparts.
Source: Materials Today Communications
Author: Saurab Singh
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: hybrid
Estimated APC paid: $2,830
OpenAlex ID: W7166643270