【作者】 赵鹏；

【导师】 程晓敏；

【作者基本信息】 华中科技大学 ， 电子科学与技术， 2024， 硕士

【摘要】 传统的冯诺伊曼架构在面对物联网时代下的海量数据显得捉襟见肘,作为新型存储技术之一的相变存储器因其与CMOS工艺兼容的优点被认为是神经形态器件的有力候选者。而相变材料因其快速结晶、熔融相变和阻值漂移的特性,使得相变电突触在电导调节的线性度、对称性及保持特性方面存在巨大挑战。此外,相变材料还具有丰富的半导体特性。而现有的相变光突触主要基于相变特性,在人工神经网络应用上较单一。针对上述问题,本文提出一种N0.92at.%-Ge₁Sb₄Te₇（以下简称NGST）相变电突触,并利用基于半导体特性的NGST相变光突触共同构建了一种同质光电储备池计算系统。主要工作如下:在电突触方面,通过材料计算和薄膜表征等手段对N掺杂Ge₁Sb₄Te₇的影响进行深入研究,提出通过N掺杂不仅引入Ge-N键（四面体局部结构）有效延缓Ge₁Sb₄Te₇的结晶过程,还具有抑制结构驰豫的效果,从而实现NGST相变电突触具有电导渐变和良好的多级电导态保持特性。在光突触方面,非晶态NGST的禁带宽度为0.52 eV,且禁带中存在着陷阱能级,能减缓光生载流子的复合,使得非晶态NGST具有显著的持续光电导效应。基于此,NGST相变光突触在430 nm波长下具有非线性光响应和短时程记忆特性。在神经形态应用方面,对NGST光突触的非线性映射的特性进行验证,并利用NGST相变光突触（作为储备池层）和NGST相变电突触（作为读出层）共同构建同质光电储备池计算系统,仿真结果表明该系统可以实现视觉手语识别任务,识别率达到99.58%。即使在15%的背景噪声水平下,识别率仍然能保持90%以上。更多还原

【Abstract】 The traditional Von Neumann architecture is unable to cope with the massive data in the era of Internet of Things.As one of the emerging storage technologies,phase change memory is considered as a strong candidate for neuromorphic devices due to its compatibility with complementary metal oxide semiconductor.However,due to the characteristics of rapid crystallization,melting phase transition and resistance drift of phase change materials,the phase change electrical synapses suffer from great challenges in the linearity,symmetry,and retention of conductance regulation.In addition,phase change materials have rich semiconductor properties.However,the existing phase change optical synapses are mainly based on phase change mechanism,which are relatively limited in the application of artificial neural networks.Therefore,this paper proposes a N0.92at.%-Ge₁Sb₄Te₇（NGST）phase change electrical synapse,and uses the NGST phase change optical synapse based on semiconductor mechanism to construct a homogeneous photoelectric reservoir computing system.The main works are as follows:In terms of electrical synapses,the influence of N-doped Ge₁Sb₄Te₇was studied by means of material calculation and thin film characterization.It is proposed that N-doped Ge₁Sb₄Te₇not only effectively delays the crystallization process of Ge₁Sb₄Te₇by introducing Ge-N bonds（local tetrahedral structure）,but also has the effect of inhibiting structural relaxation.Thus,the NGST phase change electrical synapse demonstrates the characteristics of gradual conductance and good multilevel retention.In terms of optical synapses,the band gap of amorphous NGST is 0.52 eV.There are traps in the band gap,which prolong the photogenerated carrier recombination process,so that amorphous NGST has a significant persistent photoconductivity effect.Based on this,NGST phase change optical synapse exhibits nonlinear optical response and short-term memory characteristics at 430 nm wavelength.In terms of neuromorphic applications,the nonlinear mapping characteristics of NGST optical synapse are verified,and a homogeneous photoelectric reservoir computing system is constructed by NGST phase change optical synapses（as the reservoir layer）and NGST phase change electrical synapses（as the readout layer）.Simulation shows that the system can realize visual sign language recognition task,and the recognition accuracy reaches 99.58%.Even at a background noise level of 15%,the recognition accuracy can remain above 90%.更多还原