Abstract
<jats:p>Thermal history is geometrically encoded in the atomic structure of glasses, but whether the geometric carrier of this memory is universal across glass families or system-dependent has not been tested. Here we apply a GATv2 graph neural network structural probepreviously validated on Lennard-Jones (LJ) and Kob–Andersen (KA) pair-potential glasses—to amorphous silica (SiO2), a tetrahedral network glass with fundamentally different bonding. Using published molecular dynamics configurations of 3000-atom glasses at four cooling rates spanning three decades (0.1–100 K/ps), with spatial patch extraction and stratified group k-fold crossvalidation to prevent data leakage, we find that the geometric carrier of thermal history is system-dependent. Bond-length statistics, which achieve >95% classification accuracy in pair-potential glasses, carry no discriminative signal in SiO2 (43.3 ± 8.2%, consistent with chance). Adding bond-angle features (O–Si–O tetrahedral and Si–O–Si bridging angles) restores accuracy to 96.3 ± 5.2% (AUC = 0.981) under five-fold group cross-validation. Simultaneously, Kolmogorov–Smirnov tests on all three partial pair distribution functions (Si–O, Si–Si, O–O) show no statistically significant difference between the fastest and slowest glasses (p > 0.05 for all pairs), establishing that the dominant structural carrier of thermal memory in silica is invisible to standard diffractionbased characterisation. These results demonstrate that the GATv2 architecture transfers without modification across glass families, but discovers qualitatively different geometric carriers in each: bond-length distributions in pair-potential glasses versus bond-angle correlations in network glasses.</jats:p>