| In the era of big data,people’s demand for computer computing power is gradually increasing.Brain-like intelligence inspired by brain science research results is ready to take off.Artificial neural network based on traditional complementary metal-oxide-semiconductor（CMOS）processes are difficult to further integrate,leading to storage and power consumption bottlenecks in their development.As a new type of information device with high speed,low power consumption,compatibility with CMOS process and similar functions to human brain synapses and neurons,memristors are potential candidates for building large-scale artificial neural network hardware systems and a powerful way to develop next-generation artificial intelligence chips.The thesis firstly models the memristor and introduces the window function to obtain an effective memristor simulation model,and simulates the basic characteristics of the memristor using Cadence software.Then,the structure of Pt/HfO2/BiFeO3/HfO2/TiN memristor is prepared by photolithography and magnetron sputtering technique,and its DC resistive characteristics and conductance change under pulse signal are tested.The results show that the resistive characteristics and operating mechanism of the device are consistent with the ion mobility simulation model,which confirms the reliable of the amnesia resistance model established in this thesis.Based on the memristor simulation model established in this thesis,a multi-terminal input neuronal circuit was constructed.A double spike pulse is used to simulate the pre-neuronal afferent stimulus,and the number and frequency of the pulses are adjusted to simulate the intensity of the external stimulus.By comparing whether the accumulated input signal voltage is higher than the threshold voltage of the circuit,the leaky-integrate-fire（LIF）characteristic of the neuronal signaling process is realized.Further,changing the threshold voltage of the circuit（corresponding to neurons in different parts of the organism）,it was found that when the input signal was the same,the neuronal circuit could output the voltage signal when the threshold voltage was lower;when the threshold voltage was higher,the neuronal circuit could not output the voltage signal;these results simulated the consistency of the phenomenon that different neurons were stimulated and the higher the excitation level of the neuron,the easier it was to respond.This shows that the circuit not only simulates the LIF characteristics of neuronal signals during transmission,but also simulates the response of neurons at different levels of excitation.Based on the research of neuronal circuits,the thesis further builds a circuit based on memristors to realize biological learning functions.An associative learning circuit is designed,which consists of a single memristor and CMOS device.The output of this circuit is consistent with the phenomenon of the famous Pavlovian dog conditioned reflex experiment,and successfully simulates the behavior of biological secondary learning and delayed learning.The circuit solves the problems of a single simulation situation and incompatibility of multiple functions in the bionic learning circuit.As shown above,the thesis models the memristor and investigates its ion migration characteristics.The Pt/HfO2/BiFeO3/HfO2/TiN structure device was also developed to obtain the resistive performance and multi-level conductance regulation characteristics,which confirmed the reliability of the memristor model.Further,a multi-terminal input neuron circuit was built to realize the LIF function of biological neurons in processing signals;the thesis also built an associative learning circuit to realize the functions of Pavlovian associative memory and secondary and delayed learning.The thesis research not only explores the potential applications of amnestic resistor-based bionic circuits in future artificial neural network systems,but also promotes the development of a new generation of artificial intelligence chips. |
