碲化物忆阻突触器件机制及联想学习功能研究

Memristive artificial synaptic device is the basic building block of neuromorphic chips or systems, which is of great scientific significance in developing novel brain-inspired computing architecture to deal with the massive real-time data in the Big-Data era. This new computing paradigm will break the von Neumann bottleneck of the traditional computing architecture, and facilitates the combination of data storage and computing in the chip. However, the physical mechanism is still obscure for the synaptic behaviors, and most researches focuses on the device-level cognitive functions while learning at circuit or network level is still at the very early stage. To solve these key problems, the project studies the memristive mechanisms of telluride-based artificial synaptic devices and the network-level cognitive function: associative learning. We will reveal the dependence of memristive characteristics on the transport behaviors of silver ions in the telluride films, based on which analog memristive devices with favourable repeatability will be developed. The model of the stimulus and response for the synaptic device will be established to guide the design and optimization of synaptic device and neuromorphic circuit. An associative learning circuit and methodology will be presented, and the learning with temporal contiguity and unlearning behaviors under conditioned stimulus and unconditioned stimulus will be demonstrated. Finally, we will thoroughly understand the relationship between the microscopic physical evolutionary processes and the macroscopic cognitive learning functions, and master the methodology of synaptic device design and implementation. The successful implementation of this project will promote the development of memristive synaptic devices, and also accelerate their applications in large-scale neuromorphic systems.

忆阻突触器件是神经形态芯片及系统的基本单元，对于开发面向大数据时代海量实时数据并行处理的仿脑计算架构，实现信息存储和处理的融合具有重要的科学意义。然而，现阶段突触器件的物理机制还不清晰，仿脑功能实现还停留于器件级。本项目针对此关键科学问题，重点围绕碲化物忆阻突触器件机制及联想学习认知功能展开研究。阐明碲化物材料中电场驱动下Ag金属离子输运行为对器件忆阻特性的调控规律；遴选出稳定性和重复性俱佳的渐变型忆阻材料体系；建立突触器件的信息传输与处理模型，设计构建联想学习电路网络，实现符合生物时序临近关系和预测原则的联想学习及记忆遗忘功能。理解器件宏观仿脑认知功能实现过程中的微观物理工作机制和演化过程，掌握人工突触器件的功能设计与实现方法。本项目研究将推动忆阻突触器件的优化发展，并为忆阻突触器件在大规模神经形态电路系统中的应用奠定基础。

忆阻突触器件是类脑神经形态系统的基本单元，是实现信息存储与计算融合，突破冯诺依曼计算瓶颈的核心器件。本项目重点围绕新型碲化物忆阻材料，研制出高速、低功耗的忆阻器原型，研究了器件的电阻转变物理机制、突触可塑性行为、忆阻神经网络及非易失逻辑运算等内容。本项目成功在GeTe基忆阻器中实现了长时程易化（LTP）、长时程抑制（LTD）以及时序依赖突触可塑性（STDP）等重要的突触功能，并将其作为忆阻多层感知机网络（MLP）、脉冲神经网络（SNN）中权重更新的训练法则，在MNIST手写字体库基准测试任务上达到了超过90%的识别率。此外，项目还在GeTe基忆阻器中实现了忆阻非易失状态逻辑功能，提出了基于1M1R结构的时序逻辑方法和电路。本项目的研究成果为后续发展忆阻存算一体技术奠定了器件和方法基础。研究成果参与出版专著1部（《忆阻器导论》科学出版社），在IEEE Electron device letters，Applied Physics Letters，Nanoscale，Advanced Electronic Materials，ACS Applied Materials & Interfaces等SCI期刊发表论文11篇，并多次在国际学术会议做报告，申请中国发明专利8项，其中已获授权3项。项目培养了博士2名，硕士3名。