Research Of Multi-valued Logic Model And Mechanism Based On HfOx Memristor

As the feature size of the chip decreases,mainstream memory faces severe challenges.It is an urgent need to find small size,high speed,high durability,low-power memories to replace the Flash.Resistive random access memory(RRAM)has become one of the most promising new types of memory due to its excellent performance.Hf O₂-based RRAM has received widespread attention,and this paper studies its resistance switching mechanism.The main research contents are as follows:1.Pt/Hf O₂/Ti devices with Hf O₂thickness of 8nm and 10nm were prepared by using ALD Hf O₂ film as the key dielectric material,DC scanning after forward Forming shows that the device has typical bipolar resistance characteristics,while after negative Forming,the performance of the device is very poor and prone to hard breakdown.In contrast,the 10nm device has the best overall performance.Fitting the I-V characteristic curve in different coordinate systems,it is found that the low-resistance state is the ohmic conductive mechanism,and the high-resistance state is the SCLC conductive mechanism.As the number of device scans increases,the on-state current decreases and the off-state current increases;the Set voltage increases,and the Reset voltage decreases.Under the low forming current limit,due to the influence of parasitic capacitance,the device will be formed to a low resistance state.Increasing the limiting current will increase the switching ratio,but it will increase the power consumption and cause the device to hard breakdown.2.Pt/Al:Hf O₂/Ti devices with an Al atomic concentration of 3%and 6%were prepared by inserting Al₂O₃ growth cycles during the deposition of Hf O₂.It is found that device doping will reduce the oxygen vacancy formation energy,make the formation of conductive filaments more controllable.The resistance change is changed to a gradual change,which has the potential for multi-value storage,and the conduction mechanism is changed to a Schottky emission mechanism.By changing the limit current of Set scanning,it is found that the current will always be stable at 1m A,and the self-compliance characteristic achieved.The forward and reverse cycle scanning of devices with different doping concentrations shows that the 6%Al doped devices have better uniformity and stability.Applying constant-amplitude periodic pulse voltage realizes uniform and continuous modulation of conductance during Reset and Set processes.3.An electric-thermal coupling model of RRAM based on oxygen vacancy conduction mechanism is established by CMOSOL.The models include ion mobility model,electric conduction model and Joule heating model,and the related parameters are the same as the previous devices.By solving the coefficient partial differential equations,we simulate the Reset/Set process of the device,and obtain the distribution of the oxygen vacancy concentration,temperature,electric potential,electric field and other parameters in the dielectric layer under different voltages.The effects of temperature,dielectric layer thickness,top electrode thermal conductivity,and conductive filament size on the resistance change characteristics of the device are studied,which is of guiding significance for the in-depth study of the RRAM mechanism.4.Based on the measured results of the experiment,the voltage-controlled memristor model with threshold is improved,and a memristor array is constructed.Based on the array,the design of a fully connected layer neural network is realized.The design includes a pulse generation module,a matrix module,and an activation module.MATLAB neural network hardware circuit of output processing module and other units.Based on the experimental results,we improved the voltage-controlled memristor model with threshold and constructed a memristor array.Based on this array,we realized the design of a fully connected layer neural network.The MATLAB neural network hardware circuit of the matrix module,activation module,output processing module and pulse generation module.The weight is obtained by ex-situ training,and the neural network is quantified by the 7-level conductance value modulated by the Al-doped device,and finally the quantized weight is mapped to the memristor through the differential pair.Recognition of handwritten digit data sets has an accuracy rate of 91%.