Establishment Of Mathematical Analysis Model For Resistive Memory And Study On Pattern Recognition

Resistive memory(RRAM),as a kind of non-linear electronic component,has the characteristics of continuous adjustable resistance and long-term maintenance,which are highly similar to the biological synapstic plasticity.Its structure,power consumption,density and switching speed are also comparable with synapstic,which has been considered as the best candidate for synapse emulation and pattern recognition.However,the research is limited into more high-order and complex synaptic emulation because of the undefined resistive switching mechanism and the singleness of synaptic plasticity simulation.Hense,it is urgent to explore the resistive switching mechanism,clear the law of parameter evolution and enrich the synaptic plasticity learning rules'simulation.In this article,we first construct a mathematical analysis model to explore the mechanism of the resistive switching.Then,based on the model study,we use the synaptic plasticity with time-dependent plasticity and frequency-dependent plasticity in the further pattern recognition research.The main research contents are as follows:1.A mathematical analysis model for resistive memory with different dielectric layers is used to explore the switching mechanism.In the mathematical analysis of the conductive filament mechanism,the difference in the size of the conductive filaments will affect the key parameters of the model such as the Joule heat flow and the defect migration rate.We designed a new switching model based on the defects migration under the electric field,thermal field and concentration gradient,where the model parameters are matched to the size of conductive filaments.The accuracy of the model is verified in the perovskite device.The previous models are simulated on single-layer devices.However,the switching position,switching sequence of conductive filaments will influenced by the insert layer in dual-layer device.We employed a strategy of step-by-step method to the simulation of set and reset process,and designed a model for the two-layer resistive device,which was verified in AIST/a-C devices.The simulation results of the two models are consistent with the resistive switching phenomenon,and the local temperature and morphology evolution of the conductive filaments during the switching process are obtained.2.Based on the model research,the synaptic time-dependent plasticity is simulated and their differences in performance in pattern learning tasks are explored.The use of resistive memory for synaptic plasticity simulation not only enables high-order synaptic function bionics,but also the basis for the development of brain-like neural networks.Based on the study of perovskite device model,we designed a neuron with 2T1R structure,and realized the simulation of synaptic anti-STDP through signal design.Then it is applied in single-and multi-pattern learning tasks.The coexistence of digital and analog-type switching behavior are observed in WO_x devices.Digital devices have the characteristics of discrete resistance change and lager fluctuation,while analog devices is of continuous resistance change and small flctuation.In the subsequent pattern learning tasks,they showed the advantages of fast learning speed and high learning accuracy,respectively.In order to exploit their learning advantages simultaneously,we designed a neural network with digital and analog coexistence.Through the optimized learning strategy,the network has realized the adjustability of pattern learning in terms of learning speed and accuracy.3.The pattern recognition of high-order synaptic function in visual cortex was explored by simulating the BCM learning mechanism with frequency-dependent plasticity.At present,the study of synaptic bionics mainly focuses on time-dependent plasticity(STDP)simulation.However,the BCM learning rules based on frequency-dependent plasticity are more in line with the biosynthesis plasticity rule.We are performing tri-STDP learning rule simulation on WO_x devices.Based on the all-to-all framework,we established the mapping relations between tri-STDP and BCM leraning rules.The simulated BCM learning rule is consistent with its high-frequency enhancement,low-frequency suppression,and frequency threshold sliding adjustment.Further,a 81×1 neural network is designed and applied to the BCM-based frequency pattern selective and direction selective tasks.The results show that the neural network will eventually select a certain frequency pattern or direction pattern as the training result.Frequency response on post-synaptic of the selected pattern is higher than the frequency threshold,whose synaptic weight enhanced continuously.While the post-synaptic frequency response for other patterns are lower than the frequency threshold whose synaptic weight are suppressed continuously.