Research On Multi-value Storage And In-Memory Computing Of Memristor

With the emergence of the 5G era and the big data,the fast and efficient computing systems are needed to meet the demand of information technology development.Existing computer technology is coming to a standstill due to the failure of Moore’s Law and the bottleneck of the Von Neumann architecture.The research of in-memory computing is becoming the new direction for the semiconductor industry.The memristor has the characteristics of in-memory computing,therefore it has the faster speed and the lower power consumption than the traditional computing and memory devices as for the same amount of computing task.In addition,the resistive switching materials such as ZnO and HfO₂ are compatible with the standard CMOS processes,therefore,the memristor is showing a promising future not only as a storage unit but also as an arithmetic unit.In this dissertation,the resistance switching performances of the memristor havebeen improved by the micro-nano processing method,and the logic function and neural network function have been realized by using the electrical control and circuit connection based on the memristor.The main research contents of this dissertation are summarized briefly as follows:1.We developed the universal tip-enhanced approach to modulate the performance of oxide-based memristor.When the voltage is applied to the memristor by using the scanning probe microscope tip as a microelectrode,the oxygen anions can be directionally migrated in the local area,which significantly change the morphology and oxygen vacancies distribution inside the HfO₂ nanofilm.The formation of the conical concave with controlled dimension not only reduces the thickness of the oxide layer but also induces the metal protrusion from the top electrode into the HfO_x nanofilm.Together with the increase in the local concentration of oxygen vacancies and the enhancement of the local electric field consequently resulted in the nucleation and growth of the single conductive filament,so that the integration density,stability and reliability were greatly improved for the memristor.In the pretreated memristor,the extremum ratio of the low resistance state was reduced from 10² to 10,and the extremum ratio of the high resistance state was reduced from 10⁴ to 5.The atomic point contact structure inside the memristor has been formed by controlling the reset process.The continuous modulation of sixteen quantum conductance states in the memristor has been achieved for the first time,which allows the memristor to realize the multi-value storage and the high-order neuromorphic operation.2.A neural network capable of handwritten digit recognition was constructed based on the quantum conductance of memristor.In the memristor with the structure of Pt/HfO₂/Pt,the continuous regulation of 32 conductivity states was realized by changing the reset process.The mathematical model composed of the pulse number and the conductance distribution probability was constructed through the statistical analysis of the conductance state during cyclic modulation.We also studied the effects of the conductances quantity on the recognition rate of artificial neural networks.The simulation results show that when the number of conductance state is greater than 16,the artificial neural network can achieve the handwritten digit recognition function.When the number of conductance states is 32,the artificial neural network can be trained by linearly regulating the memristor.Finally,the network has achieved the recognition accuracy close to 90%.3.We proposed the method to realize three-valued logic based on memristor.First,a bipolar three-state memristor with the columnar nano-crystalline ZnO as the switching layer was prepared.Then,the memristor was endowed with symmetrical resistance switching characteristics by setting the limiting current and the cut-off voltage.Finally,the functional complete logic set was achieved with a single memristor through the univariate operation of less than three steps.Based on this result,a three-value multiplier unit was designed by using four memristors.In addition to the in-memory computing capability,the high order computation scheme with the ternary routes may further enhance the efficiency with reduced the wire connections.By adopting the material with high dielectric constants,the memristor current can be further lowered to enhance the energy efficiency of in-memory computing.4.The adaptive machine vision system can be realized by constructing the perovskite sensor array with the memristor structure of Au/CsFAMA/ITO.The sensory memristor exhibits full sensitivity to visible spectra,as well as switchable photovoltaic and reconfigurable photoresponsivity due to the migration of ions and the interaction force among the cations in the organometallic halide perovskite.The full-function perceptron neural network can be constructed to achieve the high-fidelity imaging and in-sensor computing.Therefore,the designed perovskite photoelectric sensor array not only integrates the object imaging and recognition function in the all-in-one sensor neural network,but also realizes the adaptive adjustment of the visual imaging through the sensing layer.The perovskite photovoltaic sensory array exhibits the great potential in the smart machine vision for the vehicle autopiloting and the humanoid robot applications.