Research On Neuromorphic Devices Based On Magnetic Tunnel Junctions

Artificial Intelligence(AI)is considered as one of the core driving forces of the latest round of technological revolution.However,mainstream AI chips based on Von Neumann architecture have difficulty in improving computing power and have low energy efficiency,which hinders the development and deployment of AI technology.Therefore,it’s urgent to develop new AI chips.With brain-like architecture and extremely high energy efficiency,neuromorphic chips based on new materials and new physics is believed to be the best scheme for building new AI chips.As representative spintronic device,magnetic tunnel junctions(MTJs)have advantages of long endurance,low power consumption,nonvolatility,high speed,and abundant functions,which is considered as one of the ideal candidates for constructing neuromorphic chips.At present,the key challenge in this field is how to use MTJ to emulate different characteristics of neurons and synapses efficiently.This dissertation presents experiments on emulating some characteristics of synapes and neurons using the magnetic tunnel junctions and investigate the application of spintronic devices in neuromorphic computing by combining spintronics and artificial neural network technologies.The main contents of this dissertation are summarized as follows:1.Research on mimicking the neural population coding based on MTJ.The spintorque diode(STD)effect of an in-plane MTJ under the external magnetic field was studied.It was found that its response curve under large magnetic field was a symmetrical Lorentz curve,similar to the bell-shape tuning curve of a single neuron in the biological neuron population.Further,by applying different magnetic fields,multiple tuning curves were obtained based on the time division multiplexing of a single device.On this basis,a variety of nonlinear functions were constructed using the population coding mechanism,including the ReLU and Sigmoid activation functions,which are extensively used in AI.Then,a spintronic neural network was built to recognize images of handwritten digits in the Mixed National Institute of Standards and Technology(MNIST)database.A recognition rate of up to 94.88%was achieved.Finally,by calculating the change of the recognition rate of the neural network in the case of single neuron loss,the computational robustness of the neural network based on population coding was verified,and its advantages were discussed.2.Study on emulating low-power sparse neuron based on the STD effect of MTJs.First,MTJ with perpendicular magnetic anisotropy(PMA)was designed and fabricated.Subsequently,the STD effect with and without d.c.bias current was experimentally investigated.It was found that the microwave detection performance can be improved by two orders of magnitude when applying a d.c.bias of-20 μA and utilizing the injection-locking mechanism.With the help of a weak magnetic field,the sensitivity can be further increased to 20000 mV/mW.Moreover,the effective magnetic field of the device is partially offset by the PMA,leading to a lower d.c.bias current density,which is 105 A/cm2 and is one order lower than previously reported.In addition,the relationship between rectified voltage and the d.c.bias current was similar to the characteristics of artificial neurons with activation threshold.On this basis,a three-layer neural network with STD neurons was constructed to recognize pictures in MNIST and Fashion-MNIST database,with high accuracies of 94.92%and 86.83%and sparsity of 28%and 27%at an injection power of 0.5 μW.Finally,the neural network performance using STD neurons with non-monotonic output characteristics was also studied.3.Research on microwave-oriented neuromorphic device based on MTJ.First,the microwave synaptic device based on MTJ was studied.MTJs with all-in-plane magnetization were designed and prepared.By measuring the STD effect,it was found that the microwave response curves under a weak magnetic field is antisymmetric and meets the requirements put forward by the theoretical study on microwave synaptic devices.By changing the power of the injected microwave signal,we verified that the STD could perform multiplication computation(Vdc=Prf×Weight).The magnetic field was further used to change the resonance frequency of the devices,and the weight control of the STD synapse was realized(Weight=w(ffmr-frf)),preliminarily proving that STD can be used as microwave synaptic device.Secondly,microwave neuron based on MTJ was studied.MTJ with MgO cap layer was designed and prepared.By studying the microwave emission characteristics under the magnetic field,it was found that when the magnetic field is-300 Oe,the device can realize invariable emission frequency with the changing driven current,which is qualified for acting as microwave neuron device.Further,an artificial neural network based on spin-torque oscillator was constructed,whose recognition performance is as comparable as software neurons with a recognition rate of 92.28%.