Graphene-Ferroelectric Field Effect Transistors As Neuromorphic Devices For Brain-Inspired Computing

In this thesis,the memtransistive synapses based on graphene ferroelectric field effect transistors(GrFeFET)are fabricated and the properties are systematically investigated.Convolutional neural network(CNN)and spiking neural network(SNN)that use GrFeFET synaptic elements are constructed to execute supervised learning tasks.First,the GrFeFET fabrication method is elaborated.Then the electric transport properties of GrFeFET are systematically investigated and a physical model is developed to quantitatively explain the top-gate modulation of channel conductance.By exploiting the modulatable bipolar characteristic,the GrFeFET can be electrically reconfigured as potentiative or depressive synapse and in this way complementary synapses are realized.At last,synapse and neuron circuit is then constructed to simulate the level-based CNN and spike-based SNN respectively.Based on the latter one,remote supervised method(ReSuMe),which is a typical supervise learning algorithm of SNN,is executed and successful learning of handwritten digit images is demonstrated.The main content addresses the following issues:(1)Fabrication of the three-terminal memtransistive GrFeFET.The source,drain and top gate electrodes act as the three synaptic terminals respectively.Graphene is used as the channel,while the organic ferroelectrics of polyvinylidene fluoride(PVDF)is used as top gate dielectric for nonvolatile modulation of graphene channel.(2)Device physics of GrFeFET is systematically investigated.Tranfer curve of Si O₂back gated and PVDF top gated GrFeFET is characterized.Back and top gate modulation mechanism is explained.A physical model is built up to quantitatively analyze electrical properties of GrFeFET.(3)The mimicking of GrFeFET as three-terminal biological synapse.Graphene channel imitates the synapse,source and drain electrodes connects presynaptic neuron and postsynaptic neuron respectively.Top gate electrode receives spike signals to change synapse weight.The three-terminal synaptic memtransistor successfully demonstrate the key synaptic behavior of the analog weight modulation.Non-ideal synaptic factors are also valuated,which indicate the excellent synaptic quality of our device.(4)Graphene can be electrically modulated into hole-dominate or electron-dominate transportation.By exploiting this bipolar characteristic,the GrFeFET can be electrically reconfigured as potentiative or depressive synapse and in this way complementary synapses are realized.(5)Level-based CNN is constructed by using GrFeFET as synapses and handwritten digit images(MNIST database)are succesufully learned and recognized.In the case of CNN constructed by potentiative/depressive synapse,network-level simulation still shows good performance of 93%/94%of recognition rate even after introduce the non-ideal synaptic factors into CNN.(6)Spike-based neural network(SNN)is constructed by using GrFeFET or complementary GrFeFET pair as synaptic element with ReSuMe as weight update algorithm.3×3-pixel z,v and n images recognition task is accomplished with the above SNN.By using complementary GrFeFET pair as synaptic element,ReSuMe circuit module in SNN is much more simplified which is more favorable for hardware implementation.