A Charge-Domain Design of Ferroelectric Tunneling Junction Synapse for Spiking Neural Networks

Abstract

A charge-domain design of synaptic circuits based on ferroelectric tunnel junctions (FTJs) is proposed in this work. The device characteristics of experimental FTJs are analyzed, including their varactor properties, voltage-modulated tunneling electro-resistance (TER), and their significant displacement current, which challenge the circuit designs. A charge-domain design strategy is deployed to adapt this uniqueness, and a synaptic FTJ cell with the spiking-time-dependent-plasticity (STDP) rule is developed. With a TCAD-simulated FTJ of representative characteristics, the synaptic cell comprises seven transistors and one FTJ, mitigating the impacts of capacitive current and amplifying the differences of resistance states. Functional design and parameter tuning of leaky integrate-and-fire (LIF) neurons were performed. Using SPICE simulation, FTJ synapses were connected to LIF neurons, forming a low-power, unsupervised spiking neural network (SNN). Trained and tested on Modified National Institute of Standards and Technology (MNIST) handwritten digits, it achieved 88% classification accuracy. This confirms the self-learning ability of the FTJ-based neural network circuit, offering insights for neuromorphic computing advancements.