Nature Communications recently published a research paper titled Full-cycle device-scale simulations of memory materials with a tailored atomic-cluster-expansion potential, which has made significant advances in atomic-scale device-level simulations. This offers a valuable theoretical tool for the research and design of next-generation phase-change memory technology and neuromorphic computing chips.

The Materials Innovation Design Center (CAID) of the State Key Laboratory for Mechanical Behavior of Materials at Xi'an Jiaotong University (XJTU) collaborated with researchers from the University of Oxford to develop an ultrafast machine learning potential using the "Atomic Cluster Expansion" (ACE) framework.

This work builds upon the GST-GAP machine learning potential previously developed by the team (Nature Electronics 2023, 6, 746–754). By introducing the ACE framework and an additional domain-specific dataset iterative process, the new GST-ACE potential function further enhances simulation efficiency without sacrificing the accuracy of atomic forces and motions. It achieves over 400 times the efficiency of GST-GAP, and enables molecular dynamics simulations of millions of atoms at the nanosecond scale or billions of atoms at the picosecond scale.

With this new simulation tool, the research team was able to simulate, for the first time, the entire working cycle of a phase-change memory device at the atomic scale. The researchers not only reproduced the picosecond-level rapid amorphization process (RESET, data erasure operation), but also simulated the crystallization process (SET, data writing operation), which lasts only nanoseconds and involves complex behaviors such as random nucleation and grain growth.

The simulation examined two main device structures —cross-point and mushroom-type — and dynamically tracked the evolution of atomic structures under applied pulses. This work bridged the longstanding gap between atomic simulations and actual devices, providing atomic-level insights into key scientific questions, such as structural evolution, crystallization randomness, and multi-logic-state stability in brain-inspired computing of phase-change memory materials across multiple read-write cycles.