Recent Progress in Artificial Synapses Based on Two-Dimensional van der Waals Materials for Brain-Inspired Computing

Abstract

On the basis of recent research, brain-inspired parallel computing is considered as one of the most promising technologies for efficiently handling large amounts of informational data. In general, this type of parallel computing is called neuromorphic computing; it operates on the basis of hardware-neural-network (HW-NN) platforms consisting of numerous artificial synapses and neurons. Extensive research has been conducted to implement artificial synapses with characteristics required to ensure high-level performance of HW-NNs in terms of device density, energy efficiency, and learnings accuracy. Recently, artificial synapsesspecifically, diode- and transistor-type synapsesbased on various two-dimensional (2D) van der Waals (vdW) materials have been developed. Unique properties of such 2D vdW materials allow for notable improvements in synaptic performances in terms of learning capability, scalability, and power efficiency, thereby highlighting the feasibility of the 2D vdW synapses in improving the performance of HW-NNs. In this review, we introduce the desirable characteristics of artificial synapses required to ensure high-level performance of neural networks. Recent progress in research on artificial synapses, fabricated particularly using 2D vdW materials and heterostructures, is comprehensively discussed with respect to the weight-update mechanism, synaptic characteristics, power efficiency, and scalability.

Understanding the Memory Window Overestimation of 2D Materials Based Floating Gate Type Memory Devices by Measuring Floating Gate Voltage

“…[ 14 ] As well as FETs, 2D materials have been utilized in non‐volatile memory (NVM) devices, which are one of key building blocks for modern integrated circuits. [ 15,16 ] In particular, lots of efforts have been devoted to floating gate (FG) type memory devices. Their device structures have also evolved from the early stage of MoS 2 channel/HfO 2 tunnel barrier/multilayer graphene FG [ 17 ] stacks to recent 2D hetero‐stack of WS 2 channel/ h ‐BN tunnel barrier/graphite FG, [ 18 ] where the large memory window over 20 V for gate voltage ( V G ) range of ± 25 V and a good retention characteristic of 13% charge loss after 10 years have been demonstrated.…”

Artificial van der Waals hybrid synapse and its application to acoustic pattern recognition

“…The operating energy for reading and updating a weight were approximately from 0.12 to 0.71 nJ (for reading energy) and 0.79 (for updating energy in the potentiation channel)/0.93 (for updating energy in the depression channel) pJ, respectively, where a spike with 1 V of amplitude and 10 ms of duration was applied 45 . Such dissipated energy per event were determined by P = I × V × t duration 12,16,18 , and relevant details are provided in Supplementary Fig. 14.…”

“…Woo et al and S. Park et al reported HfO x RRAM and AlO x /TiN PCMO-based synaptic devices, respectively, where the nonlinear and asymmetrical conductance change resulted in a low inference accuracy of <40% for the Modified National Institute of Standards and Technology (MNIST) dataset 26 , 27 . For the HfO x RRAM, an AlO x barrier layer was introduced, consequently improving its conductance change more linearly, but this device inherently presented low dynamic range and high CCV 18 , 26 . For the hafnium-zirconium oxide FeFET-type synapse reported by M. Jerry et al, a highly symmetric conductance change was achieved, leading to a high inference accuracy of 90% for the MNIST dataset 28 .…”

“…Therefore, considerable effort has been directed toward the implementation of artificial synapses mimicking the behavior of biological synapses, such as short-term plasticity and long-term plasticity 6,7 . Synaptic devices based on various operating mechanisms and materials have been reported, including resistive random-access memory (RRAM), phase-change memory (PCM), field-effect transistors (FETs) with ferroelectric or charge-trapping layer, electrochemical memory, optoelectronic memory [8][9][10][11][12][13][14][15][16][17][18][19][20] . However, it has not been demonstrated that an HW-NN composed of such synaptic devices can perform training and inference tasks with the same level of accuracy as a software-based neural-network (SW-NN).…”

Bioelectronics‐Related 2D Materials Beyond Graphene: Fundamentals, Properties, and Applications

“…According to the definition of presynaptic terminals to input stimuli, the synaptic devices based on these 2D materials can be categorized as transistor‐ or memristor‐type artificial synapses. [ 222,230,231 ] The former are usually three‐terminal devices that use electric/optical gates as presynaptic input terminals and 2D semiconductors with source–drain electrodes as the postsynaptic output terminal ( Figure a,b). [ 225 ] The source–drain current after a presynaptic input signal is defined as post‐synaptic current (PSC) (Figure 10c).…”