The phase-change material, GeSbTe, is the canonical material ingredient for next-generation storage-class memory devices used in novel computing architectures, but fundamental questions remain regarding its atomic structure and physicochemical properties. Here, we introduce a machine-learning (ML)-based interatomic potential that enables large-scale atomistic simulations of liquid, amorphous, and crystalline GeSbTe with an unprecedented combination of speed and density functional theory (DFT) level of accuracy. Two applications exemplify the usefulness of such an ML-driven approach: we generate a 7200-atom structural model, hitherto inaccessible with DFT simulations, that affords new insight into the medium-range structural order and we create an ensemble of uncorrelated, smaller structures, for studies of their chemical bonding with statistical significance. Our work opens the way for new atomistic insights into the fascinating and chemically complex class of phase-change materials that are used in real nonvolatile memory devices.

Understanding the relation between the time-dependent resistance drift in the amorphous state of phase-change materials and the localised states in the band gap of the glass is crucial for the development of memory devices with increased storage density. Here a machine-learned interatomic potential is utilised to generate an ensemble of glass models of the prototypical phase-change alloy, Ge 2 Sb 2 Te 5 , to obtain reliable statistics. Hybrid density-functional theory is used to identify and characterise the geometric and electronic structures of the mid-gap states. 5-coordinated Ge atoms are the local defective bonding environments mainly responsible for these electronic states. The structural motif for the localisation of the mid-gap states is a crystalline-like atomic environment within the amorphous network. An extra electron is trapped spontaneously by these mid-gap states, creating deep traps in the band gap. The results provide significant insights that can help to rationalise the design of multi-level-storage memory devices.

Phase-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 Ge2Sb2Te5. 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(N)), 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 ps−1) 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 ps−1) is comparable to the minimum cooling rate needed in order completely to vitrify the models on quenching from the melt.

The radiation hardness of amorphous GeSbTe phase-change random-access memory material has been elucidated by ab initio molecular-dynamics simulations. Ionizing radiation events have been modeled to investigate their effect on the atomic and electronic structure of the glass. Investigation of the short- and medium-range order highlights a structural recovery of the amorphous network after exposure to the high-energy events modeled in this study. Analysis of the modeled glasses reveals specific structural rearrangements in the local atomic geometry of the glass, as well as an increase in the formation of large shortest-path rings. The electronic structure of the modeled system is not significantly affected by the ionizing radiation events, since negligible differences have been observed before and after irradiation. These results provide a detailed insight into the atomistic structure of amorphous GeSbTe after irradiation and demonstrate the radiation hardness of the glass matrix.

The nature of chemical bonding of molybdenum in high level nuclear waste glasses has been elucidated by ab initio molecular dynamics simulations. Two compositions, (SiO)-(BO)-(NaO)-(CaO)-(MoO) and (SiO)-(BO)-(NaO)-(LiO)-(MoO), were considered in order to investigate the effect of ionic and covalent components on the glass structure and the formation of the crystallisation precursors (NaMoO and CaMoO). The coordination environments of Mo cations and the corresponding bond lengths calculated from our model are in excellent agreement with experimental observations. The analysis of the first coordination shell reveals two different types of molybdenum host matrix bonds in the lithium sodium borosilicate glass. Based on the structural data and the bond valence model, we demonstrate that the Mo cation can be found in a redox state and the molybdate tetrahedron can be connected with the borosilicate network in a way that inhibits the formation of crystalline molybdates. These results significantly extend our understanding of bonding in Mo-containing nuclear waste glasses and demonstrate that tailoring the glass composition to specific heavy metal constituents can facilitate incorporation of heavy metals at high concentrations.