基于忆阻器nT1R构架实现多级阻态渐变调控的人工神经突触器件研究

It is one key step to develop artificial synapse with inherent cognitive functions for exploiting brain-analogous intelligent system. The synaptic simulation using memristor can be regarded as one ideal choice. Noted that the multi-state switching of memristor is one criterion to affect the accuracy of synaptic emulation, such as number of states, resistance switching region, gradual switching adjustment. This project would aim at developing one new operation system based nT1R (n-Transistor/1-Resistor) structure and new memristive behavior that possessing both abrupt and gradual switching for mimicking synapse. It also can achieve adjustable gradual multi-state switching with large region. Different from the familiar mode of applying bias on the two-terminals electrode, the operation mode of nT1R construction is that the gate electrode of MOS transistor and the top electrode of memristor serve as pre-synapse and post-synapse, respectively. This kind of operation mode has several advantages, such as accurate control of the conducting filament size, signal integration from multi-neurons and compatibility with high density technology, which can This project plans to study the adjustment of memristive devices based on nT1R construction; to build up artificial synapse using series and parallel connections between multi-synapses; to achieve accurate simulation for primary synaptic functions and adaptive evolution functions, and explore advanced synergetic functions based on multiple synapses; to study the applications of synaptic devices, such as pattern learning. This study can extend the implementation scheme of synapse devices and promote the precision for mimicking synapse, which has certain creativity and is one leading edge topic.

开发具有认知能力的人工神经突触是发展类脑智能系统的核心，利用忆阻器实现对突触的仿生模拟被认为是理想选择。忆阻器多级阻态的量级数目、阻变范围、连续可调性是影响突触仿生准确性的重要判据。本项目旨在拓展一种基于忆阻器nT1R（n-Transistor/1-Resistor）构架的突触仿生运行新体系,发展兼具阻态突变与渐变特性的忆阻新形式，实现阻态变化范围大且可控的多级阻态渐变调节。与常见的两端电极运行不同，基于nT1R结构的运行以晶体管栅极和忆阻器顶电极充当突触前部和突触后部。此构架具有精确调控导电通道尺寸、整合多传输信号、高密度工艺兼容等优点。本项目将基于nT1R构架开展忆阻器件调控、人工突触构建(多突触串/并联)、认知功能仿生(基本突触功能、自适应演变功能、高级协同功能)、应用探究（模式识别等）等研究工作，能够扩充神经突触仿生器件实现方案并提升器件仿生准确性，具有创新性，属于国际前沿课题。

本课题基于氧化物、碳基等忆阻材料，面向高性能忆阻器件研制及类脑突触功能实现的研究目标，围绕忆阻器导电通道多级阻态调控的基本问题，主要从高性能提升、人工突触构筑、类脑功能模拟及应用等三个方面开展了深入研究。提出了引入离子缓冲层、构筑微纳孔道和光电协同作用等新方法，实现了对导电通道的精细调控，研制了具有3000万次超长寿命、7.5fJ超低能耗的高性能忆阻器件；基于高可靠器件，提出了1-transistor/1-resistor（1T1R）、2T1R等新型人工神经突触仿生构架；实现了对大脑经典/高级认知规律的类脑模拟，创建了具有时空模式识别功能的忆阻型人工视觉系统。项目取得了如下主要成果：（1）引入离子缓冲层技术，获得了超过3000万次的超长擦写寿命碳基忆阻材料与器件；提出了构筑微纳孔道的新方法，利用其尺寸限域和界面限域效应，获得稳定有序的超低尺寸通道，据此实现了低于大脑突触100fJ能耗的7.5fJ超低能耗；（2）设计了1T1R、2T1R、数字/模拟混合型等忆阻神经突触构架，促进了导电通道的精细调控和多级阻态的连续调节，获得了准确率达95.6%的高效图像识别；（3）采用界面能带设计和电学性质调控，构建了基于Pt/WOx肖特基势垒调制的二阶忆阻器件，提出了将经典的“三脉冲时序依赖可塑性（triplet-STDP）”模型推广至BCM理论的全新模拟方案，构筑忆阻型人工神经网络；提出了利用三脉冲刺激调控频率依赖可塑性的新策略，完整模拟了大脑高级认知规律“Bienenstock-Cooper-Munro（BCM）理论”，研制了具有时空模式识别功能和高效图像识别能力的人工视觉系统；开发了等离激元型新型光电忆阻器，实现了全光信号调制的类脑视觉神经形态器件。本项目研究成果对深入理解忆阻器件多级阻态演变过程的物理机理具有重要参考价值，所提出的忆阻器仿生人工突触构架，为构建阻态高度可调的忆阻器件及其对人工神经突触进行高精度仿生研究提供了新的研究思路，研究成果可为实现人工神经网络提供理论基础。