2D Material Based Synaptic Devices for Neuromorphic Computing

Cao, Guiming; Peng, Meng; Chen, Jiangang; Liu, Haishi; Bian, Renji; Zhu, Chao; Liu, Fucai; Liu, Zheng

doi:10.1002/adfm.202005443

Cited by 187 publications

(172 citation statements)

References 263 publications

Supporting

Mentioning

168

Contrasting

Unclassified

Order By: Relevance

“…Another crucial element, particularly for neural networks online training, is the high speed of synaptic devices. [31] We tested the switching speed of the (O-ECRAM)-III device using various gate pulses with varying pulse width, as shown in Figure 4I. We observed that the (O-ECRAM)-III device shows volatile characteristics with 5 ms pulse width.…”

Section: Resultsmentioning

confidence: 97%

“…To implement the hardware for neuromorphic computing, high endurance, and good retention characteristics in artificial synapse devices are the primary concern. [31] During the training of hardware neural networks, large number of weights updating cycles are necessary to be carried out on synapse devices. With this concern, endurance up to 10 3 voltage gate pulse operations for LTP and LTD was demonstrated in the (O-ECRAM)-III device, as shown in Figure 4E.…”

Section: Resultsmentioning

confidence: 99%

See 1 more Smart Citation

All‐Solid‐State Oxygen Ion Electrochemical Random‐Access Memory for Neuromorphic Computing

Nikam

Kwak

Hwang

2021

Adv Elect Materials

View full text Add to dashboard Cite

mimicking the brain-like functionality are being explored. [2] The development of artificial synapse devices with better synaptic plasticity and high speed is the demand of present technology. Extensive integrated circuits on complementary metal-oxidessemiconductor (CMOS), [3] and two terminal memristors such as phase-change memory (PCM), [4] resistive randomaccess memory (RRAM), [5] and other filament-based memory devices have been attempted for synaptic neural functions, but they lack in symmetricity and linear switching in conductance level; also they consume massive space and energy due to large scale complexity.However, three-terminal electrochemical random-access (ECRAM) may allow for improved multi-level memory storage due to separate read and write paths. Solid-state ECRAM was first demonstrated in 1989 using hydrogen doping of WO 3 , though those failed after 25 cycles and needed high voltage. [6] Recently, ECRAM has been proposed as synaptic devices to build artificial neural networks (ANNs) for neuromorphic computing. [7] An ECRAM, also referred to as "redox transistor or synaptic transistor," consists of a mixed ion-electron conductor channel into which an ion-conducting electrolyte pumps ions under the influence of a gate electrode. [8] The operations of ECRAM are based on modulating the device conductance by ionic doping/de-doping. [6][7]9] In ECRAM, the conventional gate dielectric was substituted by an electrolyte gate, which allows the conduction of various mobile ions, e.g., H + , Li + , and O 2-. The ECRAM synapse has a two-step mechanism: firstly, the synaptic weight update is modulated by the applied gate bias; on the second part, the read is separately operated on a channel at the source-drain terminal. [10] Earlier reported Li + and H + based ECRAM shows recommended synaptic property with high endurance and low energy consumption. [7,11] Li + ion-based ECRAM (Li-ECRAM) has shown very consistent charge-discharge curves and no longer suffers from open circuit potential (OCP) issues. [12] However, Li-ECRAM suffers from a significant shortcoming, such as the devices using nonstandard materials, including lithium transition metal oxides and polymers, all of which cannot be integrated into CMOS circuits for real application. [10b,13] In other respect, proton-based ECRAM (H-ECRAM) devices exhibit a non-zero OCP that is resulting in poor data retention. When the gate is Artificial synapses based on electrochemical random-access memory (ECRAM) have emerged as an important component for neuromorphic chips because they are capable to execute simultaneous signal transmission and memory operations. However, existing ECRAM synapse surfers with compatibility and rapid memory loss issue due to highly reactive Li + and H + cationic species. Here, all-solid-state oxygen ion-based ECRAM (O-ECRAM) synapse, which shows linear weight update characteristics through multi-level nonvolatile analog conductance states is presented. Crucially, an O-ECRAM device delivering the highly stable, nonvolatile multi-leve...

show abstract

Section: Resultsmentioning

confidence: 97%

Section: Resultsmentioning

confidence: 99%

All‐Solid‐State Oxygen Ion Electrochemical Random‐Access Memory for Neuromorphic Computing

Nikam

Kwak

Hwang

2021

Adv Elect Materials

View full text Add to dashboard Cite

show abstract

“…The employment of optical modulation is suitable not only for high energy efficiency and ultrahigh computing speed, but also for visual‐perception systems 17‐22 . Recently, many studies on photonic synapses based on carbon nanotubes, 23,24 perovskites, 25‐30 nanodots, 31 and 2D materials, 32‐34 have been demonstrated for neuromorphic computing. However, the main effort presently is utilizing photonic stimulation of fixed spectrum to achieve synaptic functions.…”

Section: Introductionmentioning

confidence: 99%

Spectrum‐dependent photonic synapses based on 2D imine polymers for power‐efficient neuromorphic computing

Zhang

Shi

Wang

et al. 2021

InfoMat

View full text Add to dashboard Cite

2D polymers (2DPs) have attracted increasing interests in sensors, catalysis, and gas storage applications. Furthermore, 2DPs with unique band structure and tunable photophysical properties also have immense potential for application in photonic neuromorphic computing. Here, photonic synaptic transistors based on 2DPs as the light‐tunable charge‐trapping medium are developed for the first time. The resulted organic transistors can successfully emulate common synaptic functions, including excitatory postsynaptic current, pair‐pulse facilitation, the transition of short‐term memory to long‐term memory, and dynamic filtering. Benefitting from the high photosensitivity of the 2DP, the devices can be operated under a low operating voltage of −0.1 V, and achieve an ultralow energy consumption of ~0.29 pJ per event. In addition, the heterostructure formed between the 2DP and organic semiconductor enables spectrum‐dependent synaptic responses, which facilitates the simulation of visual learning and memory processes in distinct emotional states. The underlying mechanism of spectrum‐dependent synaptic‐like behaviors is systematically validated with in situ atomic force microscopy based electrical techniques. The spectrum‐enabled tunability of synaptic behaviors further promotes the realization of optical logic functions and associative learning. This work inspires the new application of 2DPs in photonic synapses for future neuromorphic computing.

show abstract

“…applications including post-complementary metal-oxide semiconductor (CMOS) logic, [1][2][3] memory, [4][5][6] flexible electronics, [7,8] and hardware-relevant artificial intelligence development. [9][10][11] The carrier mobility (μ) is a central parameter in characterizing electron and hole transport in a material. High mobility values enable particular promise for high-performance devices, especially for field-effect transistors (FETs).…”

Section: Introductionmentioning

confidence: 99%

Mobility Extraction in 2D Transition Metal Dichalcogenide Devices—Avoiding Contact Resistance Implicated Overestimation

Pang

Zhou

Liu

et al. 2021

Small

View full text Add to dashboard Cite

Schottky barrier (SB) transistors operate distinctly different from conventional metal‐oxide semiconductor field‐effect transistors, in a unique way that the gate impacts the carrier injection from the metal source/drain contacts into the channel region. While it has been long recognized that this can have severe implications for device characteristics in the subthreshold region, impacts of contact gating of SB in the on‐state of the devices, which affects evaluation of intrinsic channel properties, have been yet comprehensively studied. Due to the fact that contact resistance (RC) is always gate‐dependent in a typical back‐gated device structure, the traditional approach of deriving field‐effect mobility from the maximum transconductance (gm) is in principle not correct and can even overestimate the mobility. In addition, an exhibition of two different threshold voltages for the channel and the contact region leads to another layer of complexity in determining the true carrier concentration calculated from Q = COX * (VG–VTH). Through a detailed experimental analysis, the effect of different effective oxide thicknesses, distinct SB heights, and doping‐induced reductions in the SB width are carefully evaluated to gain a better understanding of their impact on important device metrics.

show abstract

2D Material Based Synaptic Devices for Neuromorphic Computing

Cited by 187 publications

References 263 publications

All‐Solid‐State Oxygen Ion Electrochemical Random‐Access Memory for Neuromorphic Computing

All‐Solid‐State Oxygen Ion Electrochemical Random‐Access Memory for Neuromorphic Computing

Spectrum‐dependent photonic synapses based on 2D imine polymers for power‐efficient neuromorphic computing

Mobility Extraction in 2D Transition Metal Dichalcogenide Devices—Avoiding Contact Resistance Implicated Overestimation

Contact Info

Product

Resources

About