Performances Of Memristive Devices Based On The Ion Migration In Oxides

In the past few decades,the computing capability of digital computers based on complementary metal oxide semiconductor（CMOS）transistors has been greatly improved.But as the Moore’s Law is coming to an end,systems based on the von Neumann architecture are increasingly difficult to meet the computing requirements of today’s data-intensive computing tasks.Memristor-based neuromorphic computing,with its ability to perform calculations where data are stored,and the advantages of convenient realization of brain-inspired neural network algorithms,is regarded as a promising alternative computing solution.However,the performance drawbacks of memristors are still a major factor restricting the application;the drawbacks include the random electrical forming process which is difficult to control,the variation in the device-to-device performance,the poor device stability,and the non-linearity of the pulse programming,etc.To optimize the performance of memristive devices,transition metal oxides that can interact electronically with protons and lithium ions are selected as the research materials in the thesis,includingα-MoO₃,SrCoO_2.5 and WO_x;several types of non-filamentary resistive switching devices are designed and prepared,including two-terminal and three-terminal devices;the device performances and the resistance switching mechanism are discussed in detail,and the potential applications of the devices are explored.The main results are as follows:（1）A pulsed laser deposition system was selected to prepare high-quality epitaxial oxide films with stoichiometric ratios to study memristors based on the migration of exogenous ions.The processes of preparing epitaxial films and ways to characterize epitaxial films were discussed in detail.It was revealed that a normal plume and the appropriate deposition temperature are crucial for the high-quality films,and the growth window is narrow.The deposition pressure needed to be coordinated with the deposition distance to find the best deposition conditions.（2）A two-terminal memristor Pt/α-MoO₃/Nb-SrTiO₃ based on the proton migration was designed,which significantly improved the device uniformity.By introducing protons into uniformα-MoO₃ film through annealing in H₂/Ar atmosphere,the destructive electroforming process was avoided,which resulted in uniform switching behavior with high yield and minimal spatial/temporal variations.In addition,synaptic functions such as short-term memory and long-term memory were successfully emulated with the device,and the image memorizing function was embodied by memorizing the letter“Y”out of three letters“X,Y,Z”in a 5×5 array.Utilizing the short-term memory ability,a reservoir computing system based on the device was constructed,which successfully classified and recognized the four letters"H,U,S,T".（3）A two-terminal memristor based on the proton migration inα-MoO₃/SrCoO_2.5oxide stack was designed,which improved the stability and the number of conduction states of the device.A synergictic switching effect was proposed in the device:When protons migrate fromα-MoO₃ to the SrCoO_2.5 lattice,both layers underwent a resistance increase,due to the reduced doping level inα-MoO₃ along with the loss of protons,and the formation of H_xSrCoO_2.5with a larger direct bandgap.While protons migrated from SrCoO_2.5 toα-MoO₃,the device resistance decreased,because of the increased proton concentration in theα-MoO₃layer and the decreased proton concentration in the SrCoO_2.5 layer.A nearly linear potentiation and depression was also realized under an appropriate pulse scheme.A two-layer backpropagation neural network based on the performance of the devices acquired an accuracy of 94.3%for the recognition of MNIST handwritten digits.（4）A WO_x-based three-terminal device based on the lithium-ion migration was prepared,which improved the linearity during pulse programming.LiSiO_x was choosed as the electrolyte material.When a positive voltage stimulus was applied to the gate electrode,lithium ions migrated into the channel material,and the channel conductance increased.When a negative voltage was applied to the gate,lithium ions were extracted from the channel material and the channel conductance decreased.The convolutional neural network constructed based on the performance of the device acquired an accuracy of 90.6%for the recognition of spoken digits.