Basic Characteristics And Application Modeling Of Hafnium Oxide Memristor

In recent years,people's way of life has relied heavily on the exponential growth in computing and storage capabilities of electronic products,and historically,shrinking device size has driven advanced computing capabilities,while also increasing power con-sumption.With the failure of Moore's Law,further reducing the size of the device has become a major problem.As a high-speed,low-power,easy-to-integrate,and CMOS(Complementary Metal–Oxide–Semiconductor)process compatible device,memristor has become a next-generation device.A candidate for a generation of high-density storage and high-performance computing.At the same time,its electronic synapse and neuromor-phic computing functions can be used as the core device of the future non-Von Neumann architecture system.Based on hafnium oxide memristor and gallium oxide stacked de-vice,this thesis explores its basic resistive characteristics,multi-value characteristics and synaptic characteristics,and complete modeling according to their properties to realize the function of biological working memory and the function of spiking neural network based on unsupervised learning.The main research contents are as follows:(1)The basic I-V characteristics of the Al/Ti N/Hf O₂/Ti N/Al device,the influence of different limiting currents on the device endurance,as well as the device retention char-acteristics and switching speed are studied.The effects of the thickness of the resistive layer,the size of the electrode,and the material of the top electrode on the resistance ratio,operating voltage and yield are also studied.The results show that the operating voltage distribution of the symmetrical hafnium oxide device is concentrated.A larger limit current will reduce the endurance of the device,while a smaller value will reduce the resistance ratio.In the case of a fixed voltage amplitude,the switching speed of the device reaches hundreds of nanoseconds.The effects of different thicknesses of hafnium oxide resistive switching layer and different top electrode materials affects both the yield,operating voltage and the resistance ratio of the device.The resistance ratio is the largest when Cu is the top electrode,while the operating voltage is the smallest when W is the top electrode.(2)Based on the single-layer resistive material device,the gallium oxide resistive layer is added to form Al/Ti N/Hf O₂/Ga₂O₃/Ti N/Al stacked device,the basic I-V char-acteristics,endurance and retention characteristics of the device are studied,and the multi-value characteristics of the stack device are mainly studied,including the number of pulses and amplitude of pulses.The results show that the operating voltage of the device increases and the resistance ratio decrease after adding gallium oxide.The multi-value characteristic of the device is obvious.The device exhibits multi-value characteristics under the pulse number modulation and pulse amplitude modulation.In the two modulation methods,12non-overlapping resistance states can be found,and they can be well maintained.(3)Based on the Al/TiN/HfO₂/Ga₂O₃/Ti N/Al stacked device,the synaptic plas-ticity and volatile conductance properties of the device are realized.The results show that the paired-pulse facilitation degree of the device is closely related to the pulse inter-val.When the pulse interval is larger,the paired-pulse facilitation phenomenon is weaker.The conductance volatility behavior of the device is close to the forgetting characteristic curve.The traditional binary memristor model is improved to realize the threshold char-acteristic and conductivity volatile characteristic model of stack device.The memristor with this characteristic is used to realize the function of biological working memory,and the simulation results are close to the real results.(4)Model and simulate three common neurons of the spiking neural network,and then build a pulse based on the LIF(Leaky Integrate-and-Fire)neuron model and the synapse with STDP(Spike-Timing-Dependent Plasticity)characteristics neural network,the network can complete unsupervised learning well,and the learning ability of the net-work is tested with the MNIST handwritten data set.At the same time,the parameters of the network are changed to simulate the fluctuation caused by the inconsistency or failure of the device during the real mapping to verify the robustness of the network.Finally,the read-write circuit,current limiting circuit and mapping circuit system of memristor are described,and the three classification is realized on the two-layer spiking neural network,with an accuracy of 88%.