Physical reservoir computing-based online learning of HfSiOx ferroelectric tunnel junction devices for image identification

Abstract

Synaptic devices for neuromorphic computing, remarkably those destined for next-generation applications, are increasingly considering ferroelectric tunnel junctions (FTJs) as highly promising candidates. Notably, the ferroelectric characteristics of HfOx are substantially improved following silicon doping. This enhancement is attributed to the smaller atomic radius of silicon compared with hafnium, which facilitates optimal lattice distortion and polarization behavior, thereby making the material suitable for ferroelectric applications. This study investigates performance variations resulting from postmetallization and postdeposition annealing. Additionally, it analyzes the influences of the utilization of lift-off compared with etching techniques during the patterning process, ultimately optimizing the performance of the TiN/HfSiOx(HSO)/Si device. The device also employs a nonfilamentary gradual resistive switching memristor to simulate the behaviors of an artificial nociceptor and synapse. The fabricated HSO-based FTJ device exhibits critical biological nociceptor characteristics, including relaxation, sensitization, recovery, non-adaptation, and threshold response. By modulating input spikes, the device effectively emulates the core functionalities of biological synapses, resulting in a diverse array of synaptic plasticity responses. Computational simulations corroborate the proficiency of the device in executing both computational and sensing tasks with high efficiency.