Computationally Efficient Machine-Learned Model for GST Phase Change Materials via Direct and Indirect Learning

TL;DR

This paper introduces a fast, accurate machine-learned potential for GST phase change materials using the Atomic Cluster Expansion framework, enabling large-scale simulations of phase transitions with reduced computational cost.

Contribution

The authors develop an ACE-based ML potential for GST that matches the accuracy of previous models but is significantly faster, and they introduce an indirect training approach to leverage larger datasets.

Findings

ACE potential achieves comparable accuracy to GAP

ACE model is orders of magnitude faster, especially with GPU acceleration

Both direct and indirect training reproduce experimental and DFT results

Abstract

Phase change materials such as Ge$_{2}$Sb$_{2}$Te$_{5}$ (GST) are ideal candidates for next-generation, non-volatile, solid-state memory due to the ability to retain binary data in the amorphous and crystal phases, and rapidly transition between these phases to write/erase information. Thus, there is wide interest in using molecular modeling to study GST. Recently, a Gaussian Approximation Potential (GAP) was trained for GST to reproduce Density Functional Theory (DFT) energies and forces at a fraction of the computational cost [Zhou et al. Nature Electronics $\mathbf{6}$, 746-754 (2023)]; however, simulations of large length and time scales are still challenging using this GAP model. Here we present a machine-learned (ML) potential for GST implemented using the Atomic Cluster Expansion (ACE) framework. This ACE potential shows comparable accuracy to the GAP potential but performs orders of magnitude faster. We train the ACE potentials both directly from DFT, as well as using a recently introduced indirect learning approach where the potential is trained instead from an intermediate ML potential, in this case, GAP. Indirect learning allows us to consider a significantly larger training set than could be generated using DFT alone. We compare the directly and indirectly learned potentials and find that both reproduce the structure and thermodynamics predicted by the GAP, and also match experimental measures of GST structure. The speed of the ACE model, particularly when using GPU acceleration, allows us to examine repeated transitions between crystal and amorphous phases in device-scale systems with only modest computational resources.