Unraveling the crystallization kinetics of Ge-rich GexTe phase change alloys with a machine-learned interatomic potential

Phase change chalcogenides such as the flagship Ge2Sb2Te5 (GST) compound are exploited in key enabling technologies such as non-volatile electronic memories and neuromorphic computing. However, the crystallization temperature (Tx) of GST is too low for applications that require data retention at high temperatures such as those in the automotive sector. In this respect, Ge-rich GeSbTe alloys recently emerged as a promising candidate for high temperature applications due to their higher Tx which is ascribed to the occurrence of Ge segregation during the crystallization of the amorphous phase [1]. The details of the crystallization process at the atomic scale are, however, largely unknown. In this contribution, we report on the molecular dynamics simulations of the crystallization process of Ge-rich GexTe binary alloys which share many properties with the ternary GeSbTe system. To this aim, we have developed a machine learning interatomic potential generated from a large database of energies and forces computed within Density Functional Theory (DFT) and fitted by the Neural Network (NN) method implemented in the DeePMD package [2]. The potential is able to reproduce with high accuracy the structural and dynamical properties of Ge2Te, GeTe and elemental Ge in the liquid, amorphous and crystalline phases. The NN potential has then been exploited in large scale simulations (20000 atoms for tens of ns) to study the crystallization kinetics of the Ge2Te amorphous alloy via phase separation into the stoichiometric GeTe compound and elemental Ge.