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

Phase change memories (PCMs), typically based on the Ge2Sb2Te5 (GST225) compound, are exploited to build artificial synapses [1] and neurons [2] of interest for neuromorphic computing. Both applications rely on the partial crystallization of the amorphous phase of GST225 which allows changing nonlinearly and continuously the electrical resistance of the memory. The crystallization kinetics is an essential functional feature for these applications. However, the experimental study of the crystallization kinetics at the operating temperatures of PCMs is difficult due to the high crystal growth velocity and nucleation rate. On the other hand, in the last decade atomistic simulations based on Density Functional Theory (DFT) shed light on the early stage of crystal nucleation and growth in GST225. However, the limitations of the DFT methods in system size and simulation time prevents addressing several important issues on the crystallization kinetics of interest for the operation of neuromorphic devices such as the effect of nanoconfinement or the grain coarsening, just to name a few. To overcome these limitations, we have developed an accurate interatomic potential for GST225 by training on a large DFT database the artificial neural network (NN) scheme implemented in the DeePMD-kit [3]. The resulting interatomic potential allows the simulation of nanometric size cells for tens of ns at the atomistic level with low computational cost.In this talk, we will present the validation of the NN interatomic potential and its application to large-scale simulations of the crystallization process in GST225. [1] Duygu Kuzum, Rakesh G. D. Jeyasingh, Byoungil Lee, and H.-S. Philip Wong Nano Letters 12, 2179 (2012).[2] T. Tuma, A. Pantazi, M. Le Gallo, A. Sebastian, and E. Eleftheriou, Nat. Nanotechnol. 11, 693 (2016).[3] H. Wang, L. Zhang, J. Han, and W. E, Comp. Phys. Commun. 228, 178 (2018)