Unraveling the Crystallization Kinetics of the Ge2Sb2Te5 Phase Change Compound with a Machine-Learned Interatomic Potential

Phase change alloys, such as the Ge 2 Sb 2 Te 5 compound, are emerging materials for in-memory computing and neuromorphic applications. These applications exploit a rapid and reversible switch between the amorphous and crystalline phases induced by Joule heating. The crystallization kinetics is, therefore, a fundamental functional feature for these applications. However, the high nucleation rate and crystal growth velocity make it hard to investigate the crystallization kinetics experimentally. Atomistic simulations could thus provide important insights on the crystallization process. Indeed, molecular dynamics (MD) simulations based on Density Functional Theory (DFT) shed light on the early stage of crystal nucleation. Yet, DFT-MD is restricted to small systems and short simulation time preventing its application to issues of relevance for the operation of the devices (e.g., crystallization in confined geometries). Machine learning techniques provide a route to overcome the limitations of DFT methods as shown in the seminal paper by Behler and Parrinello [1] who introduced a Neural Network (NN) scheme to generate interatomic potentials by fitting the DFT potential energy surface. This method was used about ten years ago to generate a NN potential for the phase change compound GeTe [2] that allowed addressing several issues such as the crystallization kinetics in the bulk and in nanowires and the aging of the amorphous phase [3]. In this talk, we report on the generation of a NN potential for the prototypical phase change compound Ge 2 Sb 2 Te 5 within the NN framework implemented in the DeePMD-kit package [4]. The NN potential allows simulating several tens of thousands of atoms for tens of ns at a modest computational cost. The validation of the potential and its application to the study of the crystallization kinetics in the supercooled liquid and overheated amorphous phases will be discussed [5]. [1] J. Behler and M. Parrinello, Phys. Rev. Lett. 98, 146401 (2007). [2] G. C. Sosso, G. Miceli, S. Caravati, J. Behler, and M. Bernasconi, Phys. Rev. B 85, 174103 (2012). [3] G. C. Sosso and M. Bernasconi, MRS Bulletin 44, 705 (2019). [4] H. Wang, L. Zhang, J. Han, and W. E, Comp. Phys. Commun. 228, 178 (2018); L. Zhang, J. Han, H. Wang, R. Car, W. E, Phys. Rev. Lett. Phys. Rev. Lett. 120, 143001 (2018). [5] O. Abou El Kheir, L. Bonati, M. Parrinello, and M. Bernasconi, arXiv:2304.03109 (2023).