Neuromorphic Computing Based On Thin-Film Transistor Devices

In the big data era,neuromorphic computing has great advantages in overcoming the von Neumann bottleneck and dealing with data-intensive computation tasks efficiently.For hardware implementation of a neuromorphic system,artificial synaptic devices with high technology maturity,stable working mechanism,rich and adjustable synaptic functions and linear and symmetric conductance tuning are urgently needed.As an artificial synaptic device,multi-terminal thin-film transistor devices have the advantages on mature technology,high stability,good reliability,parallel signal processing and transmission,and high compatibility with CMOS technology,thus showing great potential in exploring artificial neural network,developing intelligent edge computing and other new information processing fields.However,there are stills many challenges when the transistors devices are applied in neuromorphic computing as artificial synaptic devices.The realization of excitatory/inhibitory synaptic function in the same device,the development of multiterminal heterosynaptic plasticity,the acquisition of non-volatile synaptic weights and the regulation of linear multi-level conductance remain to be further optimized.In view of these key issues,the physical mechanism,design and optimization of synaptic function and pattern recognition simulation in artificial neural network are thoroughly addressed in this thesis.The main research work and innovation results are summarized as follows:Firstly,in the low-temperature polysilicon thin-film transistor,fundamental bilingual homosynaptic behaviors,including excitatory postsynaptic current,inhibitory postsynaptic current and paired pulse facilitation,are successfully emulated,based on the charge trapping mechanism under electric pulse stimulation at either top or bottom gates.Besides,the strength of excitatory and inhibitory responses induced by one gate can be dynamically modulated by the electrical biases at the other gate,indicating the realization of heteroplasticity.Furthermore,the transition between excitatory and inhibitory modes can be easily controlled by the interplay of the voltage biases at top and bottom gates.To obtain non-volatile synaptic weights,the thesis demonstrated a synaptic device based on the low-temperature polysilicon thin-film transistor with floating-gate structure.Fundamental synaptic behaviors,including excitatory postsynaptic current,inhibitory postsynaptic current and transition from short-term plasticity to long-term plasticity,are successfully emulated based on the charge trapping mechanism under electric pulse stimulation at the gate.By optimizing the operation scheme to reach constant charge injection/extraction under each pulse,non-volatile 5-bit analog conductance states with small nonlinearity,and low variation are obtained.Furthermore,the feasibility of using charge-trap synapses in a three-layer perceptron is evaluated.High-accuracy face recognition of 97.5% with good tolerance on non-idealities is demonstrated.What’s more,in order to explore the realization of the visual system,the optical signal is introduced in as the synaptic stimulus in the In Ga Zn O thin-film transistor,which can be benefit to the aspects of power comsuption,bandwidth,robustness and signal transmission speed.The excitatory synaptic functions are realized due to the photogenerated carriers and oxygen vacancy ionization,and the inhibitory synaptic functions are realized based on the mechanism of electrons trapping in the top-gate insulation layer under the positive gate pulse.Since the UV light can accelerate the electron detrapping and meanwhile the PPC phenomenon under the UV light can be erased by a positive gate voltage,the transistion between excitatory and inhibitory synaptic function is demonstrated under the interaction of optical and electrical mechanisms.In particular,the potentiation and depression characteristics are adopted as the weight update rule in a three-layer perceptron and a convolutional neural network,and the recognition accuracy for MNIST recognition task reached 88.91% and 95.99% respectively.And good tolerance to input noises is also demonstrated.In summary,the research results obtained in this thesis have laid the foundation for clarifying the physical origin of synaptic transistors,design and optimization of synaptic functions,and promoting the development and applications in the field of neuromorphic computing of synaptic transistor devices.