A Neural Network Potential for Phase Change Materials
Lehrstuhl für Theoretische Chemie, Ruhr-Universität Bochum
This project is a collaboration with Gabriele Sosso and Prof. Marco Bernasconi, Universita di Milano Bicocca, Italy
Phase change materials (PCM) based on chalcogenide alloys are presently used in digital versatile disk (DVD-RW)
and are promising materials for nonvolatile electronic memories (NVM) [1]. Both applications
rely on the reversible and fast transition (about 50 ns) between amorphous and crystalline phases.
A Phase Change non-volatile Memory (see Fig.1) is essentially a resistor of a thin film of chalcogenide material
with a low field resistance which changes by several order of magnitude across the phase change, the system
being highly conductive in the crystalline form and insulating in the amorphous phase. These two phases are
the two states of the memory. Although Ge2Sb2Te5 (GST) is the most widely studied
PCM for applications in NVM, the related binary GeTe compound has also been thoroughly investigated. In spite
of this, the structural, optical and electronic properties of amorphous GeTe and GST and mechanisms driving
the phase transitions are still matter of debate. In this context, ab initio simulations turned out to be a
reliable tool, able to shed light even onto the atomistic process that rule the amorphization of these systems ([2], see
Fig.2). However, some interesting properties of PCM would require an huge computational effort in order to be
investigated at ab initio level. Neural Network potentials (NNP, builted according to the Behler-Parrinello framework [3])
are capable of providing results of the same quality with respect the ab initio ones together with a computational
speed up of several orders of magnitude. Thus, developing such a potential for this class of materials could definitely
widen the horizons of atomistic simulation in this field.
Schematic diagram of a typical NVM based on phase change materials, from Nat. Mat. 4, 347 (2005).
Four-fold rings in a model of amorphous GST obtained by means of ab initio simulations, see for example
J. Phys. Cond. Matt. 21, 255501 (2009).
[1] A. L. Lacaita and D. J. Wonters, Phys. Stat. Sol. (a) 205, 2281 (2008)
[2] S. Caravati, M. Bernasconi, T.D. Kuhne, M. Krack and M. Parrinello, Phys. Rev. Lett. 102 205502 (2009)
[3] J. Behler and M. Parrinello, Phys. Rev. Lett. 98 146401 (2007)