Device-scale atomistic modelling of phase-change memory materials using a machine-learned interatomic potential Yuxing Zhou 1,2 , Wei Zhang 2* , En Ma 2 and Volker L. Deringer 1* 1 Department of Chemistry, Inorganic Chemistry Laboratory, University of Oxford, Oxford OX1 3QR, United Kingdom 2 Center for Alloy Innovation and Design (CAID), State Key Laboratory for Mechanical Behavior of Materials, Xi’an Jiaotong University, Xi’an 710049, China *E-mail: wzhang0@mail.xjtu.edu.cn; volker.deringer@chem.ox.ac.uk Phase-change materials (PCM) are leading candidates for next-generation memory and neuromorphic computing chips, which encode digital information via a large electric or optical contrast between the crystalline and amorphous phases of PCMs. The Ge–Sb–Te alloys have been most widely studied and used in such applications, and in particular, the compounds on the GeTe-Sb 2 Te 3 tie-line (referred to as “GST”) are now used in commercial 3D Xpoint memory products, serving as an important component to bridge the performance gap between memory and storage units, and facilitating ongoing developments in neuromorphic engineering for more complex artificial intelligence tasks. Quantum-accurate computer simulations have played a central role in understanding phase-change materials (PCMs) for advanced memory technologies. However, the drastic growth in computational cost with model system size has precluded simulations on the length scales of real devices. In this work, [1] we introduced a single, compositionally flexible machine-learning (ML) interatomic potential model that was developed using a two-step iterative training process. The accuracy of this potential has been validated via comparison with reference ab initio molecular dynamics (AIMD) simulations. We show that our model can describe the flagship GST alloys under various practical device conditions, including fully atomistic simulations of non-isothermal heating, and taking chemical disorder into account. The growth simulation of GeSb 2 Te 4 (in a cell of 1,008 atoms), can be well reproduced within one day using our model, whereas the previous AIMD simulations on the same task took several months. The superior computing efficiency of the new approach enables simulations of multiple thermal cycles and delicate operations for neuro-inspired computing. The smooth overlap of atomic positions (SOAP) analysis has been used to investigate the crystallisation and amorphisation processes, which quantified the geometrical and chemical order in comparison to the idealised GST models. [2] A device-scale capability demonstration (containing more than 500,000 atoms with a size of 40 × 20 × 20 nm l ) shows that the new ML potential can directly describe technologically relevant processes in realistic PCM-based memory products. Our work demonstrates how atomistic ML-driven simulations can help study the structural and chemical properties as well as programming mechanisms of GST in device-scale models, and thus guide design for high- performance devices. References 1. Y. Zhou, E. Ma, W. Zhang, V. L. Deringer. Preprint at https://arxiv.org/abs/2207.14228 (2022). 2. Y. Xu, Y. Zhou, X.-D. Wang, W. Zhang, E. Ma, V. L. Deringer, R. Mazzarello. Adv. Mater. 34, 2109139 (2022).
DD24
© The Author(s), 2023
Made with FlippingBook Learn more on our blog