Research On Preparation Of LiSiOx-Based Memristors And Neural Network Application

With the advent of the information technology era,the amount of data to be processed by computers is growing exponentially,and computers with traditional von Neumann architecture are facing the great challenge of high energy consumption and low efficiency.Memristors,with their dynamically adjustable resistance with current and integrated storage and computation,have received widespread attention from domestic and international researchers.However,the non-ideal characteristics and non-conformity of the devices in neuromorphic computing applications based on memristors have affected the performance of neural networks.Therefore,the preparation of memristors with excellent synaptic properties and the exploration of memristors in neural networks have become a hot research topic nowadays.Based on the above needs,this paper selects lithium silicate(LiSiOx)memristors to investigate the electrical properties and neural network applications.The details are as follows.Firstly,we optimized the LiSiOx functional layer thickness based on Pt/LiSiOx/TiN,and the device was tested by electrical performance at a thickness of 15nm with high coherence(10~3s)and large switching ratio(>100).We further optimized the initialization current limit,and the device can operate stably at±2V with a current limit of 5m A,and the switching voltages during set and reset are-0.9V and 1.1V respectively,and the device shows good stability in 100 voltage cycle scans.Secondly,five initial resistances with high consistency were set for the multi-level storage capability of the LiSiOx memristor.The amnesia resistors exhibit abrupt changes in larger initial resistances and slow changes in smaller initial resistances.The relationship between the initial resistance and the height of the Schottky barrier is demonstrated through I-V curve fitting,and the conductivity mechanism of the Pt/LiSiOx/TiN memristor is then illustrated through the current transport mechanism and the conductive filament model.Finally,the LiSiOx memristor is implemented for LTP/LTD,PPF,and STDP simulated synaptic bionic functions.In the LTP/LTD simulation,the device was able to modulate in 100 conductive states and the non-linearity of the device was reduced from8.12/5.15 to 1.31/1.43 by adjusting the initial resistance.The recognition rate reached 94.58%.