Ferroelectric Charged Domain-Wall Synapse for Neuromorphic Computing.

Abstract

Inspired by brain neural networks, integrated memory-computing devices are critical to meet the demands of big data and artificial intelligence. This work explores the quasi-continuous modulation of ferroelectric charged domain walls' conductance, which is confined in a topological quad-domain, allowing the charged domain walls to serve as neural synapses. The device mimics synaptic plasticity (long-term potentiation and depression) and shows paired impulse facilitation. In a designed ferroelectric domain-wall neural network, we demonstrate multiplicative, accumulation-additive operations between the input image and the stored response matrix, capable of image processing functions, including triclassification with 100% accuracy. In the neural network simulation, the MINST database and the Cifar-10 database achieve 98.7% and 95.1% recognition rates. The sub-nanosecond polarization switching and the ultrathin (3-5 nm) charged domain walls position them as a promising platform for advancing ultrafast and scalable synaptic devices for low-power (potentially reduced to 0.2 aJ with sub-nanosecond pulse durations) neuromorphic computing systems.