Computationally efficient machine-learned model for GST phase change materials via direct and indirect learning.

Area of Science:

• Materials Science
• Computational Materials Science
• Solid-State Physics

Background:

• Ge2Sb2Te5 (GST) is a key phase change material for non-volatile memory.
• Existing Gaussian Approximation Potential (GAP) models for GST, while accurate, face computational limitations for large-scale simulations.
• Molecular modeling is crucial for understanding GST's behavior in memory devices.

Purpose of the Study:

• To develop a faster machine-learned (ML) potential for Ge2Sb2Te5 (GST) using the Atomic Cluster Expansion (ACE) framework.
• To enable simulations of large length and time scales for GST phase transitions.
• To compare the performance and accuracy of ACE potentials trained directly from DFT versus indirectly from an intermediate GAP model.

Main Methods:

• Implementation of a machine-learned potential for GST within the Atomic Cluster Expansion (ACE) framework.
• Training of ACE potentials directly from Density Functional Theory (DFT) data.
• Training of ACE potentials indirectly from an intermediate Gaussian Approximation Potential (GAP) model.
• Utilizing graphics processing unit (GPU) acceleration for enhanced computational speed.

Main Results:

• The developed ACE potential for GST demonstrates comparable accuracy to the GAP potential.
• The ACE potential achieves significantly faster simulation speeds, orders of magnitude greater than GAP.
• Both directly and indirectly trained ACE potentials accurately reproduce GST's structure and thermodynamics, aligning with experimental data.
• The accelerated ACE model facilitates the examination of repeated phase transitions in device-scale systems.

Conclusions:

• The ACE-based ML potential provides a computationally efficient tool for studying GST.
• This advancement allows for the investigation of complex phenomena like repeated phase transitions in solid-state memory materials.
• The developed potential significantly reduces the computational resources required for large-scale simulations of GST.