Compact Model Of Memristor And Its Application In Information Processing

According to the Moore's Law,the number of transistors in a dense integrated circuit doubles about every two years.Today,the feature size of semiconductor devices has been reduced from micron to 7 nanometers,gradually approaching the limit of Moore's Law.Deep nano-sized electronic devices face many complex and intractable quantum effects,bringing unprecedented challenges to the semiconductor industry.At the same time,artificial intelligence has developed rapidly,and the artificial intelligence robot AlphaGo developed by Google's DeepMind company has attracted people's attention to deep learning.AlphaGo uses 40 threads,48 CPUs and 8 GPUs.This huge number of transistors brings a shocking level of Go,which attributes to the high-speed processing of large amounts of data,big chip area and huge power consumption.Large-scale high-performance computing has made existing electronic technologies unable to meet the low power requirements of the ever-increasing number of transistors,and there is an urgent need for a low-power device that can break through the“Von Neumann”bottleneck.Memristor is a noval electronic device with excellent characteristics:?1?Memristor has a simple structure and a small device size,and can prepare a highly integrated 3D vertical structure;?2?Memristor has multiple stable states and memristance has a large dynamic range,contributing to implement multi-valued logic operations;?3?Memristor has high endurance and resistance state retention;?4?Memristor integrates storage and computing functions,etc.These properties of memristors provide a potential solution to the dilemma of existing electronic technologies.Memristors are one of the key components in the future electronic technology.In recent years,the physical preparation and resistance switching mechanisms of memristors have arouse the attention of researchers.However,the memristor models connecting the electronics design automation platform and the preparation process have not attracted much attention.The existing physical models or circuit models of memristors either have complicated computational processes,limiting the versatility of these models,or they cannot be implemented on the electronic design automation platform,resulting in the loss of the possibility of exploring the memristors'application in the circuit design.In order to solve these problems,based on the the formation and annihilation of conductive filaments,a memristor model is put forward.The versatile and accurate compact model of memristor with equivalent resistor topology has the following advantages:1)Based on the formation and annihilation of conductive filaments,the I-V characteristics of the memristor can be accurately described.2)The memristor model has a unified mathematical expression and its mathematical form is simple.3)The versatility of the model is verified by comparing the model data with the measured data in the TiO₂/TiO_2+x memristor,Pt/TaO_x/Ta memristor and Ni/NiO/Ni memristor.4)It is easy to implement this model using the hardware description language Verilog-A,and the model can be used in electronic design automation platform.4)The memristor model can also describe the cycle to cycle variations and device to device variations,providing a solid foundation for exploring various applications of memristors.However,due to the uncontrollable factors,such as the immature preparation process,the random generation and migration of oxygen ions/vacancies,the relaxation kinetics or thermoelectric effects of metal nanoparticles,the instability of the memristor is ubiquitous and affects the electrical properties of the memristor,such as the positive and negative threshold voltage,maximum current,etc.Although some strategies are put forward to improve device reliability during fabrication,variations are still inevitable,which in turn impairs the performance of the memristor application.Some strategies are as follows,such as introducing a BEOL-compatible TiN barrier layer in the Cu/HfO₂/TiN/Ru memristor to reduce nanofilament overgrowth phenomenon or a bi-layer memristor of TiO₂ with graphene oxide and reduced graphene oxide obtaining a higher switching voltage,sharper switching and higher HRS/LRS ratio.As an important memristors'application,the performance of the memristive neural network is largely affected by the instability of the memristors.In order to explore the influence of the memristors'instability on the neural network,based on a filamentary memristors'compact circuit model,two typical machine learning methods,a feed-forward network and a data clustering,as the representatives of supervised and unsupervised learnings,are tested,following the model's four variation parameters,the variations of maximum memristances,of conductive filaments'change speeds,of initial conductive filaments'lengths,and of minimum memristances.We hope the exploration can deepen the understanding of memristor's role in machine learning and give guidelines for the design and fabrication of memristive neural networks.Although memristors have the advantages of highly integrated three-dimensional vertical structures and high-speed,parallel processing of data,memristor-based image enhancement is rare.This thesis explores the feasibility of the memristor-based image enhancement and proposes two strategies of the memristor-based image enhancement:1)Memristor-based image enhancement and 2)Image enhancement based on the memristor and the transmission map.Among them,a transmission graph is calculated based on 16pixels of four corners of the image.The method is simple in calculation and can effectively compensate for the serious distortion caused by the memristor-based image enhancement.Image enhancement based on the memristor and the transmission map makes full use of the nonlinear characteristics of the memristor,reduces a large number of numerical calculations and has excellent performance.Memristor-based image enhancement is expected to provide a very efficient way to handle high-speed real-time image processing in the future.This thesis focuses on the memristor model and memristors'application in neural network and memristor-based image enhancement.The versatile and accurate compact model of memristor is established;the influence of the memristors'instability on the neural network and memristor-based image enhancement are explored.We hope our explorations provide another way of thinking for the development of memristor in the future.