Prediction Of Phase Change Memory Resistance Based On BP Neural Network

Phase change memory(PCM)has the non-volatile characteristic that traditional DRAM does not have.Compared with DRAM,PCM is more ideal in power consumption and size,so PCM is regarded as one of the most likely non-volatile memories to replace traditional DRAM.When the internal temperature of PCM is affected by the external electric pulse and other ways,the phase change layer materials will convert between crystalline(low resistance)and amorphous(high resistance).The resistance is the key factor that affects the memory performance of PCM,so it is very important to simulate the resistance of PCM.At present,PCM resistance simulation software has many,but the efficiency is not high,often a simulation takes tens of minutes.In this paper,we mainly study the change of resistance of PCM with any size in different initial states after applying any electric pulse.According to the universal approximation principle of neural network,BP neural network can approach a mapping relationship from m dimension to n dimension.Based on this principle,PCM resistance prediction method based on BP neural network is proposed.Firstly,the paper analyzes the main factors that affect the change of PCM resistance through comparative test.The influence of three variables on the change of resistance value is analyzed for the PCM with single pulse applied to change the pulse amplitude,pulse width,length,width and height of each level of the PCM and the initial phase state of the phasechange layer respectively;for the PCM with continuous pulse applied,besides the above variables,the influence of the number of pulses on the change of resistance value is also considered.Secondly,the initial phase state of PCM phase-change layer within the specified range is obtained by the simulation program written in C++.2-3 initial resistances are selected as the initial phase state of PCM phase-change layer in the specific resistance range of each order of magnitude.Then,the parameters that affect the resistance value of PCM are input into the simulation program according to the random value within the specified quantification range,and then different initial phase states are selected for multiple simulations.The resistance value,the influence parameters and the initial resistance value of phase transition layer are recorded as the data set of neural network.Finally,according to the BP neural network,training and testing are carried out with the data set obtained from simulation.Finally,through the test of the test set,the accuracy of BP neural network to predict the resistance value after the application of single pulse and continuous pulse is 97% and94% respectively.Compared with the simulation software,the prediction time of BP neural network for PCM resistance can be ignored,and the accuracy and accuracy are guaranteed,which greatly improves the efficiency of resistance simulation.