Performance Modulation And Functional Application Of Neuromorphic Devices Based On Metal Oxides

The advent of the big data era poses a challenge to the traditional computers,as it is based on the von Neumann architecture in which the separation of storage area and calculation area will cause frequent data transportation.Out of the reason,the computer faces a series of challenges such as "storage walls" and "power walls",which makes it difficult for computers to meet the current era’s requirements for high-efficiency and low-power computing.Therefore,research on new devices and architectures which are different from von Neumann’s computing has become a hot topic at present.With the vigorous developments of the researches on artificial intelligence,neuromorphic computing based on new devices has aroused strong interest of researchers.Neuromorphic computing is a new computing system that draws on the architecture and information processing methods of the brain,and then realizes the integration of information processing and storage,which can effectively break through the von Neumann bottleneck.The oxide memristor based on the ion switching mechanism has the characteristics of simple structure,low power consumption,high-density integration,etc.,and the internal conductance changes with the change of external stimuli.It has huge application potential in neuromorphic computing,and there has been vigorous developments in the realization of artificial synapses,artificial neurons,and even artificial neural networks.However,there are still some problems and challenges in the application of oxide memristors to neuromorphic computing.For example,due to the resistance switching mechanism of the oxide memristor in which the conductive filaments are in a state of random growth and rupture,the poor performance uniformity between the device to the device accompanied with leakage current problem in the cross structure have led to the limitation of the integration in memristors.In addition,there are still shortcomings in the implementation of the technical concept of "close to the brain at the device level" for neuromorphic computing.For example,the synaptic plasticity simulated by memristor only focus on the long-term plasticity,short-term plasticity,and spike-timing-dependent plasticity;and the function of brain information transmission are only simulated on synapses and neurons,while the research on the surrounding environment of synapses and neurons only stays on the circuit simulation level,and there are still some progresses to be made at the device level.In this article,we mainly focus on oxide-based memristors,taking advantage of its simple fabricating process,facilitated control of performance,and clear resistance mechanism to construct neuromorphic devices,and conduct research on the key technologies of array integration.And then the material and structure are optimized to make its performance more in line with the requirements of neuromorphic computing.Finally,the influence of the non-linearity of the memristor on the recognition accuracy of the neural network is discussed.The details are as follows:(1)Aiming at the problems of insufficient miniaturization and low integration in the process of oxide-based memristors,a high-density,low-power neuromorphic memristor array with nano size was prepared.Under the applying of electron beam lithography technology and optimizing the preparation process,a memristor array with a cell size of less than one hundred nanometers has been prepared.And the high-speed and low-power programming of the device are also realized.The achievements proved that the oxide-based memristor can reduce the device size and increase the device integration scale while ensuring high performance.For the leakage current problem in the array,based on the Hf1-xZrxO2 material,a selector with high selectivity(>5×106),low off-state current(>pA),high switching speed(-300 ns),high stability and uniformity(>107)was prepared.(2)In response to the needs of neuromorphic computing for electronic synaptic devices in simulating synaptic plasticity,artificial synapse based on the materials with high ion mobility have been proposed and prepared to achieve complex synaptic functions.A ZrO2:Y film with high ion mobility was designed and prepared by doping 8%Y2O3 into ZrO2 and physical vapor deposition,then the crystal structure and basic physical properties of the material were characterized.Subsequently,an artificial synapse was constructed based on the material of ZrO2:Y,the effect of material thickness on the polarity of the synaptic device was studied,and the reasons for the change of switching polarity were discussed.Afterwards,while optimizing the accuracy of device weights,the long-term plasticity of synapses is realized.At the same time,the synaptic metaplasticity related to the operation history is realized based on the unipolar resistance switching characteristics of the device.This result shows that it is possible to construct electronic synapse with rich characteristics by regulating the transmission characteristics of ions in the resistive switching layer.(3)Aiming at the problem that the simulation of brain information transmission unit with the current neuromorphic devices only focus on synapses and neurons,but ignore the surrounding environment,an artificial astrocyte memristors with encapsulated structure have been proposed and prepared.Under the applying of ZrO2:Y material with high ion mobility and rich oxygen vacancies to wrap TaOx based synaptic devices,an artificial astrocyte device with high uniformity and high endurance was prepared.And the refresh operation with large stimulus can effectively regulate the growth dynamics of conductive filaments and then recover the Ⅰ-Ⅴ linearity of the astrocyte memristors.Therefore,the recover property closely related to the error correction and fault tolerance of the biological nervous system was achieved,indicating that the artificial astrocyte device has great potential in neuromorphic computing.(4)An artificial astrocyte array was prepared to construct a multilayer perceptron network,and the influence of the Ⅰ-Ⅴ linearity on the recognition accuracy of the network was investigated.The array was prepared based on the optimized artificial astrocytes,and the uniformity performance of the device in the array was studied.It was found that the operating voltage and resistance state of the devices in the array were highly consistent.Taking MNIST handwritten digit recognition as the target task,and selecting the device states before and after performance optimization as the connection weights in the network,the nonlinearity after optimization will greatly improve the recognition accuracy of the neural network from 62.98%to 94.75%.This research provides ideas and foundations for metal oxide-based memristors at a deeper level of bionics,and then promotes the development and application of neuromorphic computing.