Performance Modulation And Applications Of Memristors Based On Transition Metal Oxides

Owing to the rapid development in the artificial intelligence,the traditional computer architecture is unable to meet the requirements of information and data processing.Memristive devices are used to develop artificial neural networks for high-performance neuromorphic computing,due to their diverse dynamic behaviors analogue to biological synapses and neurons.However,sneak currents in arrays and non-ideality of device properties seriously deteriorate the performances of neural networks.On one hand,memristors with self-selective characteristics are desirable for a high-density integration without extra access devices due to suppressed sneak currents.On the other hand,great efforts should be made to improve the performances of memristors.In this work,selector-less memristors and threshold memristors are achieved and applied,based on transition metal oxides.The main results are as follows:In the aspect of self-rectifying memristor,memristive devices of Pt/Cu-Nb O_x/Ti N with a large rectification ratio(?3×10³)are realized in Cu-doped Nb O_x thin films.The large rectification ratio is due to the large Schottky barrier height at the Pt/Cu-doped Nb O_x interface,which is closely related to the decreased concentration of oxygen vacancies through Cu doping.Not only the devices simulate the synaptic function of short-term plasticity,but also effectively suppress the sneak currents.More importantly,a 73×73 crossbar array of the devices(?5 Kbits)is successfully demonstrated.In the field of nonlinear memristor,double-layer Ti N/Ta₂O₅/Nb₂O₅/Pt devices are fabricated,based on the resistive switching characteristics of Ta₂O₅ and Nb₂O₅;the effects of sputtering atmosphere and annealing process on performances of devices are investigated.The devices with both selector and memristor characteristics significantly suppress sneak currents during-2 V to 2 V.In the case of threshold memristors,stochastic volatile memristive devices of Ag/Ta₂O₅:Ag/Pt are developed through Ag doping.The stochastic and volatile switching performances of Ag/Ta₂O₅:Ag/Pt device are mainly attributed to the stochastically distributed Ag in the Ta₂O₅ layer.Based on the stochastic and volatile switching performances,dropout neuronal units with adjustable trigger probability are hardware implemented.In addition,combined with packaging process of chip devices,neuron units based on Ag/Ta₂O₅:Ag/Pt devices are successfully applied in hardware-based neural networks for recognizing handwritten letters.