Oxide Neuromorphic Transistors And Their Applications In Neural Networks

In recent years,neuromorphic computing inspired by the way the human brain works has attracted great interest.The human brain is a complex network consisting of~1011 neurons and ~1015 synapses.Neurons and synapses are basic information processing units of the human brain.Therefore,the realization of artificial neurons and artificial synapses at the device level is of great significance for constructing a lowpower neuromorphic computing system.Compared with resistive coupled networks,the capacitively coupled transistor network has lower static power consumption and a better emulation of neural functions.Amorphous metal oxide semiconductors represented by indium-gallium-zinc-oxide(IGZO)have attracted great attention in the field of thin-film transistors due to high electron mobility,low process temperature,and high uniformity.In this thesis,oxide semiconductor thin-film transistors are used to emulate neuronal and synaptic functions and simulate artificial neural networks.The main research contents are as follows:(1)Schottky oxide electric-double-layer transistors(EDLTs)are used to emulate the activation characteristics of silent synapses.A high specific electric-double-layer capacitance of 16.3 ?F/cm2(1 Hz)of chitosan solid electrolyte is obtained;Schottky oxide EDLTs with a working voltage less than 2.0 V are fabricated;The modulation of Schottky barrier height between Ag and IGZO is used to emulate the activation of silent synapses.The amplitude of excitatory post-synaptic current(EPSC)excited by a single pulse after activation is about 14 times of that before activation;Then we take advantage of ionic relaxation characteristics in the electrolyte to emulate synaptic dynamic functions such as paired-pulse facilitation and high-pass filtering.(2)Multi-gate oxide electric-double-layer neuro-transistors are used to realize neuronal dendritic spatiotemporal information processing function.First,high lateral coupling specific capacitance of 4.5 ?F/cm2(1 Hz,the distance between in-plane electrodes is 565 ?m)is obtained;Second,multi-gate lateral coupling EDLTs are fabricated,and the working voltage is less than 2.0 V.On/off ratio is greater than3.5×106 and the on-state current is greater than 55 ?A even if the gate-to-channel distance is 10000 ?m;The multi-gate structure is used to modulate synaptic dynamic functions such as paired-pulse facilitation and high-pass filtering;The spatiotemporal information processing function of neuronal dendrites is mimicked,achieving a response difference up to 111%;As an example of spatiotemporal information processing functions,a simple neural network is constructed to emulate the sound localization function.(3)Oxide floating gate transistors are fabricated to realize synaptic weight updating and simulate artificial neural networks.First,IGZO floating gate transistors with Al2O3/ITO/Al2O3 stacked gate dielectric are prepared at low temperature(<80°C),and a memory window of 7.3 V is realized;32 synaptic weight states are obtained and the ratio between the highest synaptic weight state and the lowest state is greater than650;Finally,crossbar structure is used to simulate artificial neural networks to realize supervised learning function.The recognition rate of the small image version of the handwritten digital data set is as high as 95.7%.