Research On VO₂ Based Memristor Device And Its Application In Neuromorphic Computing

After entering the era of big data,there is an obvious contradiction between the inefficiency of traditional computer architecture and the need for rapid processing of data.In order to solve this problem,a direct technical approach is to use emerging storage technologies to implement an in-memory-computing architecture,such as using memristors to build artificial neural networks similar to the human brain nervous system.There have been a lot of reports about the use of a single memristor to achieve the function of synapses such as continuous adjustment of conductance and spiking timing dependent plasticity.However,the dynamics in neurons are more complex than synapses,and the study of nanodevice-based artificial neurons is still in its infancy.In this work,an oscillating neuron circuit is constructed by using Pt/VO₂/Pt memristor,which realizes important neuronal functions such as integration of membrane potential,emission of continuous spikes and oscillation frequency adjustment,and completes the simulation of spiking neural network based on experimental parameters.The main research results are as follows:（1）At the device level,polycrystalline VO₂ thin films and Pt/VO₂/Pt memristor with cross-electrode structure were implemented by magnetron sputtering and annealing.The device exhibits excellent threshold switching characteristics after 1000DC voltage sweeps.In addition,the oxygen vacancy of the VO₂ film is reduced by the second annealing process to improve devices’uniformity.There is low cycle-to-cycle variability of 0.22%and low device-to-device variability of 3.12%in optimized Pt/VO₂/Pt devices.This is attributed to the high crystalline structure of VO₂ after annealing and the uniformity of film formation by sputtering technology.The experimental results provide a reference for the preparation of high-quality VO₂ films at wafer level.（2）At the circuit simulation level,based on the connection between the switching behavior of Pt/VO₂/Pt device and thermodynamics,the LTspice XVII software was used to construct the Pt/VO₂/Pt device model,and the electrical characteristics of the device model were basically consistent with the actual experimental results in the simulation test.The oscillating neuron circuit was simulated by the device model,which successfully simulated the emission of action potential and periodic oscillation behavior in biological neurons.The output spike oscillation frequency in this circuit can be adjusted by changing the input voltage,load resistance,and parallel capacitance,which confirms the feasibility of using Pt/VO₂/Pt device to implement artificial neuron circuits.（3）At the neuromorphic application level,based on the simulation principle of neuron circuit,the oscillating neuron circuit was built by using Pt/VO₂/Pt device,and the influence of input voltage and load resistance on the oscillation frequency of the output spike was explored in the neuron circuit.When the input voltage is too large,the neuron circuit responds at a lower frequency,which is very similar to the inhibitory response in biological neurons.Further,based on the neuron circuit parameters,a three-layer spiking neural network was constructed using Python,and the network was trained and tested online using MNIST handwritten digital dataset.After 100 epochs,the recognition rate of handwritten digital pictures of the network could reach 85.78%,which confirmed the application potential of Pt/VO₂/Pt device in spiking neural networks.