Neuromorphic Devices Based On Ferroelectric Polymers

With the development of Internet of Things,the global data continues to expand and is expected to exceed 175 ZB by 2025.Because the separation of storage and computing units greatly limits computing efficiency,traditional computers cannot meet the demand for processing rapidly growing data.Neuromorphic computation with processing in memory has the potential to solve this problem because of its characteristics of large-scale parallelism and real-time information processing.In particular,neuromorphic computing based on neuromorphic devices breaks through the problem of energy consumption during information processing and becomes one of the important scientific and technological breakthroughs in the 14th Five-year Plan.Ferroelectric materials have spontaneous polarization which can be dynamically,reversibly and nonvolatilely regulated by applied electric field.By controlling the amplitude and duration of applied electric field,many intermediate states during the analog polarization reversal process of ferroelectric domain can be accurately obtained.These characteristics of ferroelectric materials are very similar to the analog weight plasticity of biological synapses,which provides a basis for developing neuromorphic devices based on ferroelectric effect.The mechanism of ferroelectric effect avoids the joule heat in current-driven devices such as those based on resistive oxides and phase change materials,enabling lower power consumption.Besides,ferroelectric materials also have excellent endurance(greater than 10¹² cycles)and environmental stability.Especially,ferroelectric polymers are flexible and biocompatible.Thus,neuromorphic devices based on ferroelectric materials are expected to hold important advantages in terms of energy consumption and functionality.However,commercialized ferroelectric devices cannot be directly applied to neuromorphic computing.The capacitive ferroelectric memory on the market obtains the stored polarization state information by reading the polarization reversal charge.The polarization reversal during the reading process destroys the original polarization state of the device,and the subsequent voltage operation is required to restore the ferroelectric capacitor to its original state.This destructive charge-based readout mechanism is not suitable for processing in memory technology that supports neuromorphic intelligent algorithms.In order to solve the problem of destructive readout of ferroelectric polarization state,my study focuses on the research of ultra-low power neuromorphic devices and systems based on ferroelectric effect by taking advantage of the similar characteristics of ferroelectric material and biological synapses.The ferroelectric material selected in this paper is ferroelectric polymers with stable chemical properties,and the main results are as follows:1.Ferroelectric P(VDF-Tr FE)thin films were prepared by spin coating method.The stable ferroelectric properties and high homogeneity of the prepared P(VDF-Tr FE)films were verified by piezoelectric force microscopy and ferroelectric tester.By analyzing the polarization reversal process and ferroelectric hysteresis loops,it is confirmed that the polarization reversal of the P(VDF-Tr FE)film is determined by both external electric field and built-in electric field(including depolarization field and imprinting field)of the film.Based on the doping controlling of low dimensional semiconductor by ferroelectric local field effect and analog reversal dynamics of ferroelectric domain,a synaptic transistor with ferroelectric modulating low-dimensional semiconductor has been realized.Under the control of the electric field,the ferroelectric domain can be gradually switched in an analog way,and presents rich intermediate states and dynamics.The ferroelectric field transistor not only has ultra-low power consumption(lower than the operating energy consumption in biological synapses,<1 f J),but also has uniform,stable and controllable performance,which greatly reduces the randomness between cycles and devices.2.The biological learning rules of spiking time-dependent plasticity(STDP)and spiking rate-dependent plasticity(SRDP)are realized in a single ferroelectric synaptic transistor.A ferroelectric synaptic transistor neural network with associative learning has been designed.The third terminal(gate terminal)of the synaptic transistor enables the device to receive the feedback signal from the neuron in real time while performing the synaptic function.Based on the above results,we realized the self-learning enhancement of connection weights between excitatory neurons in the associative learning neural network constructed by synaptic transistor array,and demonstrated the associative learning tasks such as Pavlov's dog reflexes,plate reminding kitchen and incomplete digit recalling complete digit.The associative learning neural network based on ferroelectric synaptic transistor array not only solves the delay problem caused by multiple iterations in traditional associative learning Hopfield neural network,but also has self-learning function to adapt to real-time information processing scenarios in complex environment.3.The growth parameters by magnetron sputtering of AZO and Zn O thin films was optimized.By testing the morphology,structure of these films and electrical properties of ferroelectric synaptic transistors with these films as channel and P(VDF-Tr FE)as gate dielectric,it is found that ferroelectric synaptic transistor with Zn O film grown at80 W sputtering power as channel has the best performance.Based on the above conclusions,a ferroelectric synaptic transistor device with self-rectifying effect was developed by designing the size of the source and drain electrode,which was successfully integrated into a 10×10 passive ferroelectric synaptic array.By changing the amplitude of the applied pulse,the conductance of the devices in array can be linearly regulated.Simulation show that the passive ferroelectric synaptic array can achieve the task of letter image classification and the recognition rate reaches nearly100%.4.A self-powered photovoltaic sensor based on lead-free double perovskite and ferroelectric polymer is designed.The sensor can not only receive light signals directly,but also simulate the functions of photoreceptors and nerve cells in human visual system.The residence time of photocarriers in the space charge region is greatly extended by using the ferroelectric layer to generate an energy well in the Schottky barrier,which makes the photocurrent of the device produce very strong and efficient nonlinear coupling to the sequential optical signal stimulation.By applying the nonlinear coupling to in-sensor reservoir computing,the intelligent vision system based on the self-powered photoelectric sensor array is implemented.The visual information processing tasks of static face image classification and dynamic traffic direction discrimination are successfully demonstrated.