Performance Optimization And Application Of Ag Conductive Filament Threshold Switching Devices

High-density large-scale non-volatile memory is one of the main ways to solve the failure of Moore’s Law.Among various new non-volatile devices,memristors have the advantages of low power consumption,good scalability,high density,and friendly 3D integration,which have great application potential in the field of high-density large-scale storage.However,the realization of high-density memristive arrays is limited by the leakage current problem.The key to solve the problem lies in the development of high-performance two-terminal selectors.Ag conductive filament threshold switching devices（CFTS）show broad application prospects in selectors applications due to their advantages of large on-off ratio and low leakage current.In addition,such devices can also be used to construct compact spiking neurons,enabling high-density,energy-efficient spiking neural networks.However,at the same time,CFTS devices also have some problems such as low on-state current,low endurance,poor uniformity and slow turn off speed,which need further improvement.In this thesis,researches were carried out on the performance optimization of CFTS devices and their application in the field of selectors and artificial neurons.The main research contents and results are as follows:（1）For the problems of low on-state current and slow turn off speed of CFTS devices,a new device structure was proposed to limit Ag ions implantation through a Ti N barrier layer,combined with a superlattice-like HfAl_xO_y functional layer to reduce the stability of conductive filament.The Ag/Ti N/HfAl_xO_y/Pt threshold switching devices were successfully developed,showing high ON/OFF ratio(10¹⁰),high switching speed（50 ns/500 ns）,high on-state current（1 m A）,high threshold switching uniformity（δ/μ=7.3%）and low leakage current（1 p A）.This study provides an effective device solution for high-performance selectors.（2）For the problems of poor uniformity,limited on-state current and low endurance of CFTS devices,it is proposed to use trilayered hafnium oxides to optimize the overall performance of the device.The Ag/Ti N/Hf O_x/Hf O_y/Hf O_x/Pt threshold switching devices were developed,achieving a high on-state current of 3 m A and the dispersion of threshold voltage is only 5.4%.In addition,the devices possess excellent on/off ratio(>10¹⁰),leakage current（<1 p A）,switching speed（60 ns/500 ns）,endurance（>10⁶）,and switching slope（0.0127 m V/dec）.Through multiple control experiments,XPS characterization,and first-principles calculations,this study believes that the excellent performance of the device is mainly originated from the adsorbed oxygen present in the Hf O_y layer,which will oxidize the Ag,leading to lots of Ag ions on the conductive filament in the layer.After the voltage is removed,these Ag ions will repel each other,which reduces the stability of the conductive filament and improves the on-state current and eudurance.At the same time,the location where the conductive filaments rupture is limited to the Hf O_y layer,which optimizes the uniformity of the device.Combined with the Ti N layer,which can confine Ag ions implantation,and the Hf O_x layer,which can reduce the leakage current,the device achieves excellent overall performance.（3）Threshold switching devices can also be used to realize the neuron units in brain-like computing.Two types of low-power LIF neurons were constructed based on the Ag/Ti N/HfAl_xO_y/Pt devices,namely a single device neuron and a RC circuit-based neuron.The former used the kinetic process of the conductive filaments growth in a single Ag/Ti N/HfAl_xO_y/Pt device to simulate the function of LIF neuron.By limiting the current through series resistance,the power consumption of the neuron to generate a pulse can be reduced to 16 f J.The latter utilized capacitance to integrate the input signal,and realized the basic characteristics of LIF neuron,such as threshold-driven spiking and refractory period.The neuron also achieves high energy efficiency due to the extremely low leakage current and small threshold voltage of the ATHP device.Combining LIF neuron based on RC circuits with memristive synaptic devices,a two-layer spiking neural network was constructed,and the task of handwritten digit recognition was simulated.The effects of the number of iterations and expected time on the network were studied,and a recognition rate of 87.5%was achieved.By adjusting the integral capacitance C_M,the neuron can be combined with synaptic devices with different STDP time windows to construct SNNs,and the recognition rates of 86.1%,85.6%and 84.5%were achieved,respectively.