Research On Modeling Of Temperature Coefficient Of Memristor And Its Influence On Neuromorphic Computing

Artificial intelligence has important applications in image and speech recognition,finance,medicine,games and other fields.However,the artificial neural network based on the CMOS computing hardware architecture has the disadvantages of high energy consumption and low computing efficiency when processing related tasks due to the existence of the"von Neumann bottleneck".The in-memory computing architecture based on memristor can realize the integration of computing and storage and can also efficiently complete the vector matrix multiplication operation,thereby significantly improving the computing speed and energy efficiency of the system.However,because memristors still have many non-ideal characteristics,such as inconsistency between devices,limited durability,poor analog switching capability,and crosstalk in the array,it is still very challenging for the current RRAM-based neuromorphic computing system to achieve performance equivalent to that of a software-trained neural network.Thermal stability is critical for neuromorphic terminal devices that work at different temperatures.At present,Hf O₂-based memristor is widely used due to its excellent performance.Thus,the main content of this article focuses on the temperature characteristics of Hf O₂-based memristor.First,the temperature coefficient（T_α）distribution in the Hf O₂-based memristor array is studied,and a compact model describing the statistical distribution of the temperature coefficient is established by using the two-dimensional atomic simulation method.The compact model demonstrates the excellent consistency between experimental data and simulations.The experimental test results show that the performance index of the+T_αcells is generally worse than that of the-T_αcells.Therefore,the correlation mechanism between the T_αand the retention characteristics and the analog switching performance is further studied through this compact model.Finally,the physical mechanism of T_αredistribution is explained.Based on the research on the T_αof the device and array level,the influence of the temperature coefficient distribution on the inference accuracy at different temperatures in the artificial neural network is further evaluated.A current compensation scheme and hybrid optimization method are proposed to reduce the impact of T_αdistribution.The simulation results show that after using the proposed hybrid optimization method,the recognition accuracy of MNIST handwritten digit recognition tasks using memristor based artificial neural networks at 400 K has been increased from 79.8%to 91.3%.