Non-volatile memories are nowadays employed in a wide series of applications ranging from memory cards used in several electronic devices (cameras, mobile phones ...), mass storage devices (USB sticks, MP3 players, SSDs ...), and to embedded systems or even automotive applications. Traditional non-volatile memory implementations are based on the flash technology. The 0/1 bit information is stored via trapping/removal of electrons in an electrically isolated part of the device. Their presence or absence opens or closes a conducting channel between two metal contacts in a transistor-like setup. Readout operation is performed via conductivity measure between such metal contacts. Due to the high demand for larger and larger capacity, NAND-type flash memory is the most aggressively scaled technology among electronic devices. However, the feature size of flash memory cells is close to reach its intrinsic minimum limit for scaling to higher capacity.
Phase-change materials based on chalcogenide alloys are promising candidates for flash memory replacement as next-generation non-volatile electronic memories. A phase-change memory is essentially a resistor formed by a thin chalcogenide film with a low field resistance which changes by several orders of magnitude, depending on the state of the chalcogenide, metallic in the crystalline form and insulating in the amorphous phase. The large difference in conductivity between the two states is the feature which enables information storage. Together with better scalability/higher memory density, phase-change memories offer a wide set of advantages over flash memories: higher write bandwidth, reduced power consumption, smaller latency and better endurance.
However, in spite of the great technological importance of this class of materials, the microscopic origin of several of their properties is still matter of debate. On the other hand, the performance and reliability of the device can be optimized by an accurate selection of the active chalcogenide compound. For instance, a change of several tens of degrees of the crystallization temperature can be achieved by tuning the composition of the ternary GeSbTe alloys. This dramatic change of physical properties offers wide opportunities to tailor the memory performance to specific applications and/or to improve its scalability. However, this optimization procedure cannot be left to mere trial-and-error, but requires a detailed physical understanding of the material properties. In this respect, atomistic simulations can provide crucial insights to aid the experimental activity, as for instance elucidating the correlation between functional and structural properties of these materials at the atomistic level.
Although fully ab-initio atomistic simulations have provided crucial insights on the properties of phase change materials in the lat five years, several key issues such as the thermal conductivity at the nanoscale, the crystallization dynamics, and the properties of the crystalline/amorphous interface, just to name a few, are presently beyond the reach of fully ab-initio simulations.
The development of reliable classical interatomic potentials is a possible route to overcome the limitations in system size and time scale of ab-initio molecular dynamics. Traditional approaches based on the fitting of simple functional forms for the interatomic potentials turned out to be unfeasible due to the complexity and variability of the chemical bonding in the crystal and amorphous phases revealed by the ab-initio simulations.
A possible solution is the development of empirical interatomic potentials with close to ab-initio accuracy by fitting large ab-initio databases within a neural network (NN) scheme to allow simulating thousands of atoms for tens of ns. This method has been applied successfully to study elemental sodium, carbon and silicon in the last few years. By means of this approach, we plan to devise NN potentials for phase change materials which will allow us to address the study of the crystallization dynamics, thermal transport and the amorphous/crystalline interface and their dependence on composition. We are confident that the insights provided by the atomistic simulation will aid the experimental search for better performing materials in this class for non volatile memories applications.