Recently, a research paper titled 'Machine Learning Potentials for Phase-Change Memory Materials Based on Atomic Cluster Expansion and Their Full-Cycle Device-Scale Simulation' was published in Nature Communications. This collaborative study was undertaken by the Material Innovation and Design Center of the National Key Laboratory for Mechanical Behavior of Materials at Xi'an Jiaotong University and the University of Oxford. Dr. Zhou Yuxing from the University of Oxford served as the first author, while Professor Zhang Wei from Xi'an Jiaotong University and Professor Volker L. Deringer from the University of Oxford were the corresponding authors.

Addressing the longstanding challenge of simulating germanium-antimony-tellurium (GST) alloys in commercial phase-change memory devices, the researchers developed an ultrafast machine learning potential function named GST-ACE. This function is grounded in the 'Atomic Cluster Expansion' (ACE) framework. Compared to the previously developed GST-GAP potential function, GST-ACE significantly boosts simulation efficiency by over 400 times while maintaining precise atomic force and motion accuracy. It facilitates molecular dynamics simulations at the nanosecond scale for millions of atoms or at the picosecond scale for billions of atoms.

For the first time, the research team achieved atomic-level simulation of the entire operational cycle of phase-change memory devices. This encompasses the picosecond-scale rapid amorphization process (RESET) and the tens-of-nanoseconds-scale crystallization process (SET), which involves random nucleation and grain growth. The simulations cover both cross-point and mushroom-type device structures.

This groundbreaking achievement effectively bridges the 'scale gap' between atomic simulations and real-world devices. It provides an atomic-level theoretical framework for investigating the structural evolution, crystallization randomness, and multi-logical state stability of phase-change memory materials during multiple read-write cycles and in brain-like computing applications. The research data has been made openly accessible, with the relevant potential functions and simulation data available on the Zenodo platform.