Ti/AlOx/TaOx/Pt模拟突触： memristor神经网络的关键器件

34 浏览量更新于2024-08-26 收藏 1.38MB PDF 举报

本文档聚焦于"用于忆阻神经网络的Ti / AlOx / TaOx / Pt模拟突触"这一前沿课题，发表在2018年9月的IEEE Electron Device Letters上。忆阻器（memristor）作为新兴的非volatile存储元件，因其在神经形态计算（neuromorphic computing）中的潜力而备受瞩目。作者 Yi Sun、Hui Xu、Chao Wang 等人提出了一个基于Ti、AlOx、TaOx和Pt材料构成的忆阻器作为模拟突触，应用于可塑性神经网络中。该模拟突触的核心特点是其高度的均匀性，能够展现出出色的模拟开关行为，即使在三角波脉冲下也能实现高达200种电阻状态。这使得这些突触能够在神经网络中模拟神经元之间的复杂连接，并实现精细的权重调节。长期记忆保持性能也非常显著，每个状态可以稳定维持30,000秒以上，这对于实现长时间记忆和学习至关重要。更重要的是，通过50个周期的精细编程，设备的电阻状态可以实现高精度的调整，其变化范围具有1.7%的容忍度，这为构建精确的神经网络模型提供了关键支持。忆阻器模拟突触的这些特性使其成为神经形态计算系统中模拟神经网络连接的理想选择，尤其是考虑到其高集成度、大规模并行处理能力以及对故障容错的优良性能。文章的关键词包括“神经形态计算”，“忆阻器”，“模拟突触”以及“长期保留”。整体而言，这篇研究论文为忆阻器在神经网络中的实际应用提供了新的设计思路和技术支撑，对于推动神经形态计算领域的进展具有重要意义。通过深入理解并利用这种新型模拟突触，科学家们有望构建出更为复杂且功能强大的人工智能系统。

1298 IEEE ELECTRON DEVICE LETTERS, VOL. 39, NO. 9, SEPTEMBER 2018

ATi/AlO

/TaO

/Pt Analog Synapse for

Memristive Neural Network

Yi Sun , Hui Xu, Chao Wang, Bing Song , Haijun Liu, Qi Liu , Sen Liu, and Qingjiang Li

Abstract

— Electronicsynapsewithpreciseanalogweight

tuning ability and long-term retention is the vital device

foundation of memristor-based neuromorphic computing

systems. In this letter, we propose a Ti/AlO

/TaO

/Pt mem-

ristor as an analog synapse for memristive neural network

applications. The device shows high uniformity, excellent

analog switching behaviors (up to 200 resistance states

under triangle pulses) and excellent long-term retention of

eachstate (up to 30 000 s). Furthermore, the precise modula-

tion of the device resistance state (with 1.7% tolerance) can

also be achieved by a ﬁner writing program within 50 cycles.

Index Terms

— Neuromorphic computing, memristor,

analog synapse, long-term retention.

I. INTRODUCTION

EUROMORPHIC computing based on memristive neural

networks has attracted signiﬁcant interest due to its

high scalability, massive parallelism and the excellent abil-

ity of fault tolerance [1]–[4]. Recently, various applications,

such as pattern classiﬁcation [2], [5], [6], sparse coding [7],

face recognition [8] and analog signal processing [9] have

already been implemented on memristive synapse arrays. The

core computational operation of these applications is the

reconﬁgurable analog vector matrix multiplication (VMM),

which can be implemented in memristor crossbars based on

Ohm’s law and Kirchhoff’s current law [2], [9]. Previous

studies demonstrate that memristors can successfully emulate

the synaptic plasticity (e.g. pair-pulse facilitation, short-term

plasticity, spike timing dependent plasticity and spike rating

dependent plasticity) [10]–[16]. Nonetheless, little attention

has been paid to enhancing device analog characteristics

and long-term retention property, which are actually more

Manuscript received June 16, 2018; revised July 17, 2018; accepted

July 23, 2018. Date of publication July 26, 2018; date of current version

August 23, 2018. This work was supported in part by the National Nat-

ural Science Foundation of China under Grants 61604177, 61471377,

and 61704191 and in part by the National University of Defense Tech-

nology under Grant JC15-04-02. The review of this letter was arranged

by Editor D. Ha.

(Corresponding authors: Sen Liu; Qingjiang Li.)

Y. Sun, H. Xu, B. Song, H. Liu, S. Liu, and Q. Li are with the Col-

lege of Electronic Science, National University of Defense Technology,

Changsha 410073, China (e-mail: liusen@nudt.edu.cn; qingjiangli@

nudt.edu.cn).

C. Wang is with the Key Laboratory of Nano Devices and Applications,

Suzhou Institute of Nano-Tech and Nano-Bionics, Chinese Academy of

Sciences, Suzhou 215123, China.

Q. Liu is with the Key Laboratory of Microelectronics Device and

Integrated Technology, Institute of Microelectronics, Chinese Academy

of Sciences, Beijing 100029, China.

Color versions of one or more of the ﬁgures in this letter are available

online at http://ieeexplore.ieee.org.

Digital Object Identiﬁer 10.1109/LED.2018.2860053

important to achieve valid VMM computation. The analog

switching feature provides the potential to achieve continuous

device states, and the long-term retention is the basis of

read stability and reproducibility, which are all critical to the

accuracy of VMM outputs [9], [17]. Thus, the realization of

excellent analog characteristics and long-term retention prop-

erty is crucial for current neuromorphic computing systems.

In addition, applications like off-line classiﬁcation require the

precise modulation of the synaptic weight in the crossbar

arrays [7]–[9], [17], so the precise tuning ability of the

memristive synapse is also a pivotal factor.

In this letter, we present a Ti/AlO

/TaO

/Pt based analog

synapse. Initially, the quasi-DC sweeping results demonstrate

the good uniformity and excellent multilevel characteristics of

the device. The analog tuning ability is further demonstrated

by employing positive (SET) and negative (RESET) pulse

trains with different amplitude and width parameters. Then,

retention test results show the device states can last up

to 30000s, indicating the great long-term retention property.

At last, we prove that the proposed device can be precisely

modulated to the desired arbitrary states (within the error

tolerance) by a designed writing program.

II. E

XPERIMENTS

The device has a crossbar structure with 5 μm × 5 μm

resistive area. The Pt (25 nm) bottom electrode was ﬁrstly

deposited by electron beam evaporation after the ﬁrst litho-

graphy process. Then the bi-oxide layers TaO

(5 nm) and

AlO

(3 nm) acting as resistive switching (RS) layers were

respectively deposited by electron beam evaporation after

the second lithography process. At last, Ti (30 nm) top

electrode was grown by magnetron sputtering after the third

lithography process, and an Au layer (30 nm) was covered on

Ti layer to avoid oxidation after contact with air.

All the electrical characterization was implemented with

the Keithley 4200s semiconductor parameter analyser. During

the measurements, the voltage was applied on Ti electrode

while the Pt electrode was tied to ground. The transmission

electron microscope (TEM) sample was fabricated by focused

ion beam (FIB), and the thickness in Z direction was reduced

to ∼40 nm to obtain high quality image.

III. R

ESULTS AND DISCUSSION

Fig. 1a is the SEM image and cross sectional TEM image

of the Ti/AlO

/TaO

/Pt device. The resistive area of the

See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

Authorized licensed use limited to: National Univ of Defense Tech. Downloaded on April 21,2020 at 10:17:48 UTC from IEEE Xplore. Restrictions apply.

下载后可阅读完整内容，剩余3页未读，立即下载

weixin_38693311

粉丝: 4
资源: 922

Ti/AlOx/TaOx/Pt模拟突触： memristor神经网络的关键器件

了解Ti / AlOx / TaOx / Pt模拟忆阻器的导通和开关机制

探索Ti/AlOx/TaOx/Pt模拟忆阻器的导通与开关机理

Film Stress Influence on Nb/Al-AlOx/Nb Josephson Junctions

ALOx-开源

AlOx层钝化的非晶InGaZnO薄膜晶体管的热稳定性

alox：GPU加速，分布式，Actor模型语言（WIP）

不同厚度溅射AlOx钝化的非晶InGaZnO薄膜晶体管的光照稳定性

Alox_and_Renzo_Repo:我们被诅咒的想法的地方

自对准法制备Nb/AlOx/Nb超导Josephson结的低温特性研究

AlOx钝化非晶InGaZnO薄膜晶体管的高温稳定研究

最新资源