The new generation of non volatile memories for data storage is based on a unique property of systems known as phase change materials, i.e. the super fast (ns) and reversible phase transition between the disordered and the crystalline phases. One of the reasons of the fast crystallization of GeTe-based phase change materials is the very high atomic mobility of the supercooled liquid phase, even close to the glass transition temperature. This feature is in turn a consequence of the fact that GeTe is a fragile liquid, i.e. it shows a breakdown of the Stokes-Einstein relation (SER) that relates viscosity and diffusion in the hydrodynamic regime.
In this work, we investigate by large scale molecular dynamics simulations the microscopic origin of the breakdown of the SER in GeTe. To this end we employed an interatomic potential based onto a Neural Network framework that allows to overcome the limitations of conventional first principles calculations in terms of system size and timescale.
Our findings demonstrate that the breakdown of the SER is due to the presence of dynamical heterogeneities in the atomic motion. We quantified as a function of temperature the spatial extent of domains of slow/fast moving particles. The most mobile particles tend to cluster in domains that contain a significant number of chains of homopolar Ge-Ge bonds. The number and length of these chains increases with supercooling, boosting the atomic mobility and then the fast crystallization of GeTe-based phase change materials. We also found a certain degree of cooperative motion in this system, which is due to both first-shell correlated and string-like motion. Finally, we investigated the role of the domains of most immobile and mobile particles during crystallization, finding that mobile particles tend not to crystallize, but instead to flow around immobile (crystalline) nuclei facilitating the atomic rearrangement at the liquid-crystal interface.