Research On The Memristor Based On Single-crystalline LiNbO₃ Thin Film And Its Application In Near-sensor Reservoir Computing

In sensor signal processing systems based on Von Neumann computing architecture,the problems of energy consumption and response delay caused by Von Neumann bottleneck make this information processing mode gradually unable to keep up with the demands of sensor data processing in the Internet of Things era.Inspired by biological nervous system,neuromorphic computing systems based on emerging non-volatile memory devices（also known as memristors）,that is,through memristors connected to sensors develop artificial sensory systems,which can simulate the manner of information processing in biological nervous system,providing a opens a novel avenue for processing massive sensor data.It is considered as a new computing paradigm and hardware platform to break the Von Neumann bottleneck.In order to meet above development requirements,on the one hand,the memristor should have reliable resistive switching characteristics with a low operating voltage to be directly driven by sensor signals,and memristor should be insensitive to background information（such as noises）,but can effectively distinguish the target sensor signals.On the other hand,memristor should have abundant conductance states and a considerable dynamic range of multiple weight updating to fully map the spatiotemporal characteristics of sensor input signals and realize neuromorphic computing.Therefore,an analog-type LiNbO₃（LN）memristor based on crystal-ion-slicing（CIS）and Ar⁺ion irradiation was carried out in this dissertation,which aims to develop memristive neural networks for processing the sensor signals with neuromorphic computing paradigm.The main works are as follows:（1）A highly-quality single-crystalline LiNbO₃ thin film was prepared by CIS technique and used as the memristive layer to eliminate the adverse effects of stochastic defects on the device stability and uniformity.On this basis,a uniform defect layer richen in oxygen vacancies（V_os）was introduced on the surface of single-crystalline LiNbO₃ thin film via Ar⁺ion irradiation forming a unique memristive bilayer of amorphous LiNbO₃interface layer（i.e.,defect layer）and single-crystalline LiNbO₃ bulk layer,which guarantees the LiNbO₃ memristor has reliable resistance characteristics.The LiNbO₃memristor achieves excellent data retention characteristics,very low cycle-to-cycle variance（2.15%）,device-to-device variance（4.06%）,and stable endurance property for only 0.22%fluctuations during 500000 switching cycles.（2）LiNbO₃ memristor possesses abundant synaptic characteristics and multi-weight updating performance.The interplay of such amorphous LiNbO₃ interface layer（i.e.,defect layer）and single-crystalline LiNbO₃ bulk layer on the resistive switching mechanism of LiNbO₃ memristor was revealed through temperature-dependent I-V measurements in a wide temperature range.The interface layer provides a V_o reservoir for V_o migration,and the highly ordered oxygen octahedral structure of the single-crystalline LiNbO₃ bulk layer provides stable trap sites for electron Mott variable-range hopping（Mott-VRH）transport.Under the programming of voltage pulses,the LiNbO₃memristor exhibited short-and long-term synaptic plasticity and bionic memory properties.In addition,the dynamic range,number of intermediate states,linearity,symmetry,endurance（repeated switching between intermediate states）,cyclic variances of intermediate states,repeatability,and multi-state retention capacity of LiNbO₃memristor were comprehensively analyzed.An extremely large dynamic range（up to1891.78）with tens of thousands of intermediate states and outstanding linearity and symmetry of weight update were realized in the proposed LiNbO₃ memristor.（3）An infrared near-sensor reservoir computing（RC）system based on LiNbO₃memristors and LiTaO₃ pyroelectric infrared sensors was designed and developed.The application potential of LiNbO₃ memristor in fully-connected neural networks was verified by a multi-layer perceptron to recognize the MNIST handwriting dataset.The separation property of LiNbO₃ memristor was studied.It has been proved that LiNbO₃memristor can distinguish target sensor signals effectively and is insensitive to background information,such as noises.Without signal filters,ADCs,cache,and other auxiliary modules,the infrared near-sensor RC system realized infrared sensor signal processing with spatiotemporal feature fusion.The large dynamic range with abundant conductance states of LiNbO₃ memristor guaranteed sufficient margin for infrared sensor signals to distinguish each other and map to high-dimensional space.The dynamic gesture recognition and prediction task was successfully demonstrated with little training resources,including a small dataset and training times.