A Neuromorphic FTJ Model-Driven Design of Charge-Domain Synaptic Circuits and Spiking Neural Networks

Spike Neural Networks (SNNs) operate in an event-driven manner and have excellent hardware compatibility, resulting in lower power consumption compared to Artificial Neural Networks (ANNs) that use pure software algorithms. In this work, the Spike Timing Dependent Plasticity (STDP) characteristics of Ferroelectric Tunnel Junctions (FTJs) were simulated. We used FTJ as an electronic synapse and proposed an SNN circuit that does not directly read the tunneling current, thereby avoiding the influence of FTJ’s self-capacitance on the conduction current. The functionality was verified in a small-scale array. Using the STDP data, the SNN algorithm was applied for image classification on the MNIST dataset, achieving an accuracy rate of 89.63%. This provides a promising solution for the full hardware implementation of SNNs based on FTJs.