jats:titleAbstract</jats:title> jats:pPhase-change memory materials are promising candidates for beyond-silicon, next-generation non-volatile-memory and neuromorphic-computing devices; the canonical such material is the chalcogenide semiconductor alloy Gejats:sub2</jats:sub>Sbjats:sub2</jats:sub>Tejats:sub5</jats:sub>. Here, we describe the results of an analysis of glassy molecular-dynamics models of this material, as generated using a newly developed, linear-scaling (O(jats:italicN</jats:italic>)), machine-learned, Gaussian approximation potential. We investigate the behaviour of the glassy models as a function of different quench rates (varied by two orders of magnitude, down to 1 K psjats:sup−1</jats:sup>) and model sizes (varied by two orders of magnitude, up to 24 300 atoms). It is found that the lowest quench rate studied (1 K psjats:sup−1</jats:sup>) is comparable to the minimum cooling rate needed in order completely to vitrify the models on quenching from the melt.</jats:p>