A Study On Device And Integration Technologies Of Memristor For Artificial Neural Network Applications

In recent years,with the rapid development of information technologies such as big data,cloud computing,and artificial intelligence,the global data has put on a spurt.Instead of being compute-centric,the information processing has gradually transgressed to a data-centric paradigm.However,faced with the problem of“von Neuman bottleneck",it is very difficult to meet the performance requirement for information processing in big data era.In order to fundamentally solve the "von Neuman bottleneck",the most useful way is to break away from the traditional von Neuman architecture and realize "computing in memory".The human brain,with its massive parallelism,adaptive and self-learning capabilities,and an extremely low power consumption,has no clear boundary between data storage and processing.Inspired by the human brain,building a highly efficient neuromorphic computing system that computes in memory has attracted more and more attentions.Therefore,researchers are aiming at the study of new nano-devices,hoping to realize the emulation of synapses and neurons in human brain.Memristor,also known as Resistive Random-Access Memory(RRAM),is a two-terminal device whose conductance can be continuously modulated by electric stimuli.It has been considered as ideal synaptic emulator due to its superior performance such as high speed and low power consumption.In addition,artificial neural networks implemented by memristive crossbar arrays are expected to realize highly efficient bioinspired computing systems in hardware.However,there are many challenges to be resolved at both the device level and array level in realizing memristive neural network.In this article,we mainly discuss the linearity problem at device level and leakage current problem at array level,which will give guideline for the optimization of memristive devices and the realization of crossbar array integration.The main works are as follows:(1)In the application of neural network recognition,the I-V linearity and the conductance change linearity will greatly influence the recognition accuracy.The improvement of conductance change linearity has been widely studied.In this work,we studied the I-V linearity and conductance change linearity in TaOx based memristor with different electrodes.To further explore the origin of the linearity differences,this work analyzed the resistance change mechanism by measuring the temperature dependencies of device conductance.The results show that devices with composition modulation mechanism have better I-V linearity and conductance change linearity than devices with gap modulation mechanism.Finally,based on the linearity data,this work performed neural network simulation on the MNIST handwritten datasets,thereby futher illustrating that selecting an appropriate electrode to ensure the device to have good I-V linearity and conductance change linearity will help improve the recognition accuracy of the neural network.(2)The crossbar array suffers unavoidable cross-talk problem due to the leakage current flowing through unselected cell,which will cause misreading problem,affect the power consumption of the array and limit the array size.Therefore,in order to solve the leakage current problem,a selector device is needed.In this work,we designed and fabricated a TaOx-based selector with excellent performances,which have the advantages of low leakage current,self-limiting current,and CMOS compatibility.To confirm the suppression of the selector device on the undesired sneak current,a proper memristive device was integrated on top of the selector device.After adding the selector device,the leakage current was greatly reduced.Through the calculation of the read margin and array size,the addition of selector device can greatly increase the storage capacity of the array.This work realizes a stable selector device and its integrated unit with memristor,which has certain significance in the realization of the memristive crossbar array.