Efficient Training of Machine Learning Potentials by a Randomized Atomic-System Generator.

Area of Science:

• Materials Science
• Computational Chemistry
• Condensed Matter Physics

Background:

• Machine learning potentials (MLPs) offer accurate and efficient simulations for large-scale systems.
• Developing cost-effective training datasets that represent the potential energy surface is crucial for MLPs.

Purpose of the Study:

• To develop a novel scheme for generating comprehensive training datasets for MLPs.
• To apply this scheme to create MLPs for chalcogen-based phase change materials.

Main Methods:

• The randomized atomic-system generator (RAG) scheme combines random sampling and structural optimization.
• MLPs were constructed using RAG-generated training sets for binary GeTe.
• Molecular dynamics simulations were performed to analyze melting and crystallization processes.

Main Results:

• The RAG-generated training set effectively covers the potential energy surface, including amorphous phases.
• MLPs accurately simulated the melting and crystallization dynamics of GeTe, comparable to first-principles methods.
• Phonon density of states were calculated from simulations to analyze vibrational properties during crystallization.

Conclusions:

• The RAG scheme provides a cost-effective method for generating high-quality training data for MLPs.
• The developed MLPs enable accurate simulations of phase change materials, advancing materials discovery and design.