Online and offline learning using fading memory functions in HfSiOx-based ferroelectric tunnel junctions

Abstract

Ferroelectric tunnel junctions (FTJs) are garnering significant attention as leading candidates for next-generation synaptic devices in neuromorphic computing. In particular, HfO_x-based FTJs offer several advantages over perovskite-based FTJs, including complementary metal–oxide semiconductor compatibility, scalability, low power consumption, and rapid operation. Furthermore, the ferroelectric properties of HfO_x are enhanced through silicon doping, as silicon's smaller atomic radius than hafnium makes it ideal for ferroelectric devices. In this study, we investigate HfSiO_x-based FTJs, demonstrating robustness to thermal variation by maintaining consistent use of Si doping and materials while optimizing annealing temperatures. We also examine the high tunneling efficiency, address the inherent depolarization field challenges, and delve into their potential applications in neuromorphic computing. Specifically, this approach emulates key aspects of human brain learning, including Pavlov's dog experiments, potentiation, depression, paired-pulse facilitation, and reservoir computing (RC). To demonstrate the device capability in the information processing of neuromorphic systems, image recognition simulations are performed using the MNIST database. These include online learning mechanisms related to the outcomes of potentiation and depression, as well as offline learning mechanisms related to the results of RC.