Resistive switching in ferroelectric synaptic weights

ResistiveThe Von Neumann architecture has determined the characteristics and working principles of classical computing paradigms since 1945. CMOS technology had a major role in the improvement of modern computers thanks to advancements in device scaling. However, recent years saw the advancement in deep learning algorithms built on brain-inspired networks of artificial neurons, also referred to as artificial neural networks (ANNs), con- currently with an exponential increase of processed data, representing a big challenge for conventional computer hardware. The "bottleneck" of the Von Neumann architecture is due to the unavoidable movement of data between the central processing unit (CPU) and memory, causing latency and high energy consumption. Neuromorphic computing allows the hardware implementation of the "in-memory" computing paradigm taking ad- vantage of cross-point technologies where the data can be processed and stored within the same site. This is possible thanks to beyond-CMOS devices known as memristors, emerging devices whose programmable conductance (or resistance) states can be used to encode synaptic weights for implementing ANNs algorithms such as deep neural net- works (DNNs) and spiking neural networks (SNNs). Among these, ferroelectric Tunnel Junctions (FTJs) are a great candidate for the hardware realization of SNNs. In this work, FTJ synaptic weights built with two different ferroelectric materials - HfZrO4 and BiFeO3 - are discussed and compared. First, the fabrication processes of crossbar arrays based on the two distinct FTJs are presented. Then, particular attention is given to the BiFeO3 crossbar array; electrical characterizations of the fully processed devices show a lower On/Off ratio than what is observed in the simplified test structures on the blanket film. To this aim, temperature-dependent measurements have been analyzed in order to study the conduction mechanisms across the device, and an investigation of the resistive switching mechanism is reported. The oxidation of Ni during the processing affects the polarity of the FTJ and the On/Off ratio, which becomes comparable to that of CMOS-compatible HfZrO4 junctions. Finally, the comparison between the two materials doesn’t show any relevant advantage of one material over the other. switching in ferroelectric synaptic weights