Research On Neuromorphic Devices Based On Spintronics

Neuromorphic devices have great potential for building artificial intelligence chips.Spintronic devices have the advantages of high-density integration,low power consumption,and sufficient nonlinear dynamics,which make them have great applications in constructing high efficiency,low power neuromorphic computing systems.At present,relevant research has just started,and the main topic of this research field is exploring the ways to mimic synaptic and neural characteristics with spintronic devices and construct artificial neural networks to realize high performance,low power neuromorphic computation.This dissertation presents experimental investigations of mimicking synapses and neurons with magnetic tunnel junctions and unveils the applications for neuromorphic computing by connecting spintronics and neural network technologies.The main contents of this dissertation are as follows:(1)Research on mimicking the synaptic characteristics based on magnetic domain wall motion in the free layer of a magnetic tunnel junction.The magnetoelectric transport characteristics induced by the domain wall motion in the free layer of magnetic tunnel junctions with different structures were experimentally and theoretically studied.On this basis,a width-changed structure device was designed and fabricated,and six different resistance states have been observed by pinning the domain wall at different positions defined by the width-changes of the structure.The resistance states can be modulated with the application of an external magnetic field or a d.c.current.The experimental results are well explained by micromagnetic simulation.The results suggest that our design is expected to have applications in magnetic memory and neuromorphic systems.(2)Research on mimicking the artificial neurons based on stochastic switching of magnetic tunnel junctions.The stochastic switching phenomenon driven by the spin-polarized current in the in-plane magnetic tunnel junction was studied,and simple Bayesian inference was realized.Moreover,the influence of the perpendicular magnetic anisotropy on the energy barrier of a magnetic tunnel junction was studied.Based on this,we propose a voltage-controlled spintronic stochastic device based on a magnetic tunnel junction by introducing perpendicular anisotropy into the free layer.An ultra-low power(<1 nW)stochastic behavior was observed.The relationship between the probability of stochastic switching and the magnetic field has been theoretically and experimentally studied,showing the possibility to mimic the artificial neurons.Based on these stochastic spintronic neurons,a spintronic neural network was constructed,and the recognition of handwritten digits was realized with a recognition rate reaching 95%.Furthermore,the stochastic behavior can be modulated by a bias voltage owing to the voltage-controlled magnetic anisotropy effect.The voltage-controlled stochastic behavior is theoretically and experimentally studied,which indicates that it can be used to perform as an adaptive neuron.These results provide a way for building energy-efficient spintronic neuromorphic computing systems.(3)Research on the simulation of sparse neuron characteristics based on the spin-torque diode effect in magnetic tunnel junctions.The rectification characteristics of spin-torque diodes have been investigated in the absence and presence of d.c.bias currents.While the injection locking phenomenon was observed in our devices,the output functions versus the d.c.bias currents could mimic artificial neurons with sparse representations.Furthermore,we constructed a neural network with STD neurons to recognize the handwritten digits in the Mixed National Institute of Standards and Technology database,with a produced accuracy of up to 92.7%at an injection power of 0.32?W.Further research found that the neural network based on spin-torque diode has good sparsity and stability,while its sparsity is 62%.Finally,the application of spin-torque diode neurons with non-monotonic output characteristics in neural network calculation was discussed.By optimizing the selection of network parameters,we realized the recognition of handwritten digital images with the recognition rate up to 90%by selecting random numbers from the normal distribution with the variance of 0.01 to initialize the neural network.These results suggest that STDs have potential to be building blocks for the realization of a biologically plausible neuromorphic computing system.