Simulation of Phase‐Change‐Memory and Thermoelectric Materials using Machine‐Learned Interatomic Potentials: Sb₂Te₃

Abstract

Density‐functional‐theory (DFT)‐based, ab initio molecular dynamics (AIMD) simulations of amorphous materials generally suffer from three computer‐resource‐related limitations due to their O(N3) cubic scaling with model system size, N. They are limited to a maximum model size of N ≈500 atoms; they are limited to time scales <1 ns; and, usually, only a single model can be simulated in any one investigation. This article discusses a machine‐learned, linear‐scaling (O(N)), DFT‐accurate interatomic potential (a Gaussian approximation potential, GAP), originally developed by Mocanu et al. [J. Phys. Chem. B 2018, 122, 8998] using a Gaussian process regression method for the ternary phase‐change‐memory material Ge2Sb2Te5 (GST). The chemical transferability of this GAP potential is explored in an application to the case of simulating amorphous models of the phase‐change‐memory and thermoelectric material Sb2Te3, an end‐member of the GST compositional tie‐line GeTe–Sb2Te3. The GAP‐model results are compared with those obtained from conventional DFT‐based AIMD simulations.

Gaussian Process Regression for Materials and Molecules

“…Recent studies were concerned with the supercooled liquid phase as described by GAP-driven MD, 299 assessed by comparison with experimental data from ref ( 300 ) and with the application of the Ge 2 Sb 2 Te 5 model to the end member of the quasi-binary line, Sb 2 Te 3 . 301 The latter work is a case study in transferability: studying liquid and amorphous Sb 2 Te 3 takes the potential away from the region of configuration space for which it was initially fitted. It is emphasized that the reference database for the potential contained liquid and amorphous Ge 2 Sb 2 Te 5 , in which the local environments of Sb atoms are expected to partially resemble those in Sb 2 Te 3 because of the chemical relationship between the phases, but they will be different in detail (especially beyond the first neighbor shell).…”

“…Recent studies were concerned with the supercooled liquid phase as described by GAP-driven MD, assessed by comparison with experimental data from ref and with the application of the Ge 2 Sb 2 Te 5 model to the end member of the quasi-binary line, Sb 2 Te 3 . The latter work is a case study in transferability: studying liquid and amorphous Sb 2 Te 3 takes the potential away from the region of configuration space for which it was initially fitted.…”

Inherent electron and hole trapping in amorphous phase-change memory materials: Ge2Sb2Te5

“…Bond-angle distributions typical of distorted octahedral-like atomic geometries ( i.e. a strong maximum at ∼90° and a smaller peak in the region around ∼170°) have been reported in melt-and-quenched amorphous models of phase-change materials generated by DFT-MD simulations for GST-225 30,35,46–48 and Sb 2 Te 3 , 49,50 a neural-network interatomic potential for GeTe, 51 and GAP-MD simulations for GST-225 34 and Sb 2 Te 3 , 50 as well as in a quench-rate and size-dependent study for GST-225. 52…”

A Deep Dive into Machine Learning Density Functional Theory for Materials Science and Chemistry

“…Information on the atomic neighborhoods is encoded by a bispectrum representation of the atomic density. A range of publications have employed GAPs to model the PES [370], [371], [372], [373], [374], [375], [376], [377], [378], [379], [380], [381], [382], [383], [384], [385], [386].…”