Computer simulations can play a central role in the understanding of phase-change materials and the development of advanced memory technologies. However, direct quantum-mechanical simulations are limited to simplified models containing a few hundred or thousand atoms. Here we report a machine-learning-based potential model that is trained using quantum-mechanical data and can be used to simulate a range of germanium–antimony–tellurium compositions—typical phase-change materials—under realistic device conditions. The speed of our model enables atomistic simulations of multiple thermal cycles and delicate operations for neuro-inspired computing, specifically cumulative SET and iterative RESET. A device-scale (40 nm * 20 nm * 20 nm) model containing over half a million atoms shows that our machine-learning approach can directly describe technologically relevant processes in memory devices based on phase-change materials.

    This work was done by Yuxing Zhou under the guidance of Prof. Wei Zhang, Prof. En Ma and Prof. Volker L. Deringer. This paper is free for downloads: Nature Electronics, 6, 746-754 (2023). A research briefing article was published in the same issue, highlighting the achievements made in this work: Nature Electronics 6, 726-727 (2023).

    Article link: https://www.nature.com/articles/s41928-023-01030-x

    Research Briefing: https://www.nature.com/articles/s41928-023-01031-w