Performance Optimization Of Hafnium Oxide Memristor And Its Application In Neuromorphic Computing

The explosive growth of data volume makes the transmission bandwidth between CPU and memory encounter the so-called"Von Neumann bottleneck".The development of new devices and computing architectures in the post-Moore's-law era,exploring mechanisms and methods for integrating computing and storage,and breaking through the bottleneck of chip computing power,has become an inevitable trend in the current international microelectronic integrated circuit field.Among the emerging next-generation non-volatile memories,memristors show excellent characteristics such as high speed,low power consumption,and three-dimensional integration.They can also simulate the human brain to process information quickly,parallel,and intelligent.The memristor,which mimics the brain-like neuromorphic computing method,is expected to break through traditional technology's bottleneck.Among various memristor materials,hafnium oxide is a kind of high-K dielectric material,compatible with mainstream complementary metal-oxide-semiconductor(CMOS)processes.It has excellent properties such as stable-controllable material properties and abundant physical effects,making HfO_xstand out in the technical field and have broad application prospects.However,the uniformity,speed,and power consumption restrict its commercial application.This thesis focuses on the performance optimization of hafnium oxide memristor,from the intrinsic resistance switching mechanism,device design,fabrication process,and its application in neurological calculations.The content is as follows:(1)Study on the intrinsic resistance switching characteristic of HfO_xmemristor.The Pt/HfO_x/Pt memristor was prepared by reactive sputtering at room temperature.The memristor's unipolar resistance switching was induced by the thermal diffusivity,driven by Joule heat and concentration gradient.The formation and break of the conductive filament(CF)were random,which resulted in a large drift of electrical parameters.After thermal annealing treatment at 85?and 300?for 30 min in the vacuum,the resistance switching polarity changed to bipolar.We further improved the stability performance,attributed to the uniform distribution of the oxygen vacancies—the set process's driving force mainly dominated by the electric field after thermal annealing.(2)Study on Ti/HfO_x/Pt's performance optimization by the electrode and thermal annealing process.The device fabricated at room temperature exhibits bipolar resistance switching characteristics,which has durable anti-fatigue property compared with the Pt/HfO_x/Pt device.After thermal annealing treatment at 300?for 30 min in the vacuum,the forming voltage decreased,and electrical parameters'uniformity improved.At the same time,a self-compliance and multi-level resistance capability memristor was achieved.These excellent performances are attributed to the formation of a thicker TiO_x layer after heat treatment.On the one hand,TiO_x acts as a series resistor and plays the role of divided voltage.On the other hand,the formation/rupture of the CF localized at the TiO_x/HfO_x interface,and the obtained device can show better performance.(3)Study on the theoretical feasibility of the formation of conical CF predicted by first principles.The GAA-1/2 pseudopotential method was used to perform the first-principle calculation on a 2×2×2 monoclinic HfO₂ unit cell,proving that the existence of oxygen vacancies would reduce the migration barrier of the surrounding oxygen vacancies and the operating voltage theoretically.Simultaneously,it proved that oxygen vacancies have a specific aggregation effect,which provides theoretical support for constructing oxygen concentration gradient memristors and lower operating voltage under high-speed switching conditions.A conductive cone filament induced by the oxygen concentration gradient model was proposed.(4)Study on the oxygen concentration gradient HfO_x memristor.A highly reliable and multifunctional memristor demonstrating both digital and analog characteristics was fabricated utilizing oxygen concentration gradient HfO_x films.Significantly uniform and low operation voltage(magnitude less than 0.9 V)at high speed(down to 20 ns)was realized in the digital switching mode.The performance was achieved by confining the switching region to the cone-shaped conductive channel's weak apex,which stems from the gradient distribution of resistivity and electric field in HfO_x.For ARS,the gradual resistance change is mainly attributed to the Pt/HfO_x interface barrier and the CF micromorphology modulation.(5)Study on the synaptic plasticity of oxygen concentration gradient HfO_x and its application in neuromorphic computing.Over 40 and 80 distinct levels of resistance states were defined by negative stop voltage/compliance current in DC mode.The gradually changing conductance like synapse can be tuned by pulse series,the influence of different pulse width and amplitude on the weight update of LTP and LTD process,and the trend of nonlinear factors are studied.A three-layer perceptron and CNN based on concentration memristor synapses are designed.In the three-layer perceptron simulation,the influence of different parameters(learning rate,number of iterations,number of hidden layer nodes,and noise)on the recognition rate is studied.The recognition rate can reach as high as 93.85%in the CNN network simulation.Consider the non-ideal factors.The recognition rates reached86.54%and 97.59%with the weight update rules in without-write-verify and write-verify,respectively,because write-verify can achieve higher weight update accuracy to have a higher recognition rate.