基于钙铁石氧化物相变的模拟型阻变存储器件研究

With the fast development of emerging technologies such as artificial intelligence, internet of things, etc, there have been higher and higher demands for future smart chips in terms of computing power, speed, and energy efficiency. Amid the slowdown of Moore’s law, emerging memory-based computing-in-memory (CIM) architecture has become one of the most promising technologies to break the so-called “von Neumann bottleneck” in order to improve the chip speed and energy efficiency. However, compared to conventional digital memory application, the CIM application has more stringent requirements on the performance of emerging memories. Considering the fact that current RRAM devices have shown unpleasant performance due to the random distribution and movements of oxygen vacancies, here we propose to study the Brownmillerite oxide SrFeO2.5 that has an ordered channel of oxygen vacancies in the crystal structure. The film is epitaxially grown by pulsed laser deposition, during which the direction of oxygen vacancy channel can be controlled by the growth direction. The ordered ion movement then can be modulated by external electric field, during which the phase transition is accompanied by resistive switching. Based on this physical mechanism, different types of high-performance analog resistive switching memory will be fabricated. An accurate device model will be established based on the electrical testing results and the analysis of underlying switching mechanism. Neural network simulations will be performed based on the experimental data, and methods for further device optimization will be provided. We believe that the research proposed here will provide solid device foundations for high-performance neuromorphic computing and CIM applications in the future.

在人工智能、物联网等新兴技术飞速发展的背景下，未来对于智能芯片的算力、速度和能效提出了越来越高的要求。随着摩尔定律的放缓，基于新型存储器的存算一体架构成为提高芯片算力与能效、突破“冯·诺伊曼”瓶颈的最有潜力的技术路线之一。相比于传统数字存储的应用，存算一体化的应用对新型存储器的性能提出了更高的要求。针对传统RRAM器件氧空位分布和迁移的随机性导致器件一致性差等问题，本项目选择具有原子有序的氧空位通道的钙铁石氧化物作为主要研究对象，采用脉冲激光沉积方法外延生长高质量薄膜，通过选择晶相的手段来控制氧空位通道的排布和方向，使用电场调控材料中离子的有序迁移，使其发生结构可逆相变时实现不同阻态的翻转，进而制备多种高性能模拟型阻变存储器，基于电学测试结果分析器件的阻变机理，建立准确的器件模型，并结合实验数据进行神经网络的仿真，提出器件的优化方法，为未来高性能类脑计算与存算一体应用提供良好的器件基础。

随着摩尔定律的放缓，基于阻变存储器件等神经形态器件的类脑计算是突破冯诺依曼瓶颈、实现高算力高能效芯片最具潜力的技术之一。本项目针对基于非晶氧化物的传统阻变器件存在离散性大、线性度差等瓶颈问题，提出并研制了基于拓扑相变原理的神经形态器件TPT-RAM，展现出较好的模拟阻变特性，可实现多达100个中间阻态（等效6比特以上），非线性度低至0.2和1.29，具有较低的工作电压和较快的工作速度（±1 V和100 ns），显著降低了循环间波动性（C2C）~0.9%和器件间波动性（D2D）~4.9%，室温下数据保持时间超过10年，循环耐久性超过10^8次；提出了多层次协同建模方法，结合第一性原理计算、蒙特卡洛模拟与器件的电学特性深入研究阻变机制，证实了TPT-RAM中氧空位通道的取向是影响阻变过程的关键因素；进一步利用TPT-RAM器件在不同幅值电压脉冲激励下表现出的“易失性-非易失性”阻变的双模态工作机制，首次提出了一种模拟人脑发育过程的神经网络自动剪枝策略，网络仿真结果表明TPT-RAM器件的双模态工作机制能够帮助减少82.5%的冗余突触，并将MNIST识别准确率提高到99%；在此基础上，进一步挖掘阻变器件的内在的丰富动态特性在树突网络、储备池计算等仿生神经网络方面的应用，首次构建了端到端全模拟计算的忆阻器储备池系统。该系统演示了异常心律检测、动态手势识别等时序信号处理任务，展现出了极低的系统功耗，比传统的数字计算系统低3个数量级以上。. 本项目的研究工作涵盖了阻变器件材料结构、器件工艺、机理模型、应用演示等，为后续大规模阵列集成、实现高效类脑计算打下坚实的器件基础。圆满完成项目预定的全部目标，在Science Advances、Nature Electronics、Nature Communications、IEDM等国际顶级期刊和会议上发表论文19篇，申请发明专利4项，培养博士生7名，硕士生6名。研究成果得到了国内外的广泛关注，取得了一定的国际影响力，项目负责人应邀在ICICDT、EEMD、CSTIC、MRS等国际会议上做邀请报告8次，撰写类脑计算领域的图书章节1章。