Fig. 7: Computing with oscillatory dynamics and neural networks.

a Schematic of a spin-torque nano-oscillator based on spin-transfer torque MTJ. The d.c. current injection can cause oscillation of magnetization, which results in the oscillation of MTJ voltage. b Measured a.c. voltage out of device a as a function of time, where the amplitude of the oscillation is \(\widetilde{V}\). c \(\widetilde{V}\) as a function of the injected d.c. current, where the nonlinear behavior mimics the neuron. d Schematic of a MTJ as a synapse, where the weight is tuned by the d.c. current (magnetic field). Inset is a TEM image of a MTJ. e Rectified d.c. voltage as a function of frequency of the input RF signal. f Output rectified d.c. voltage as a function of the input RF power for different synaptic weights. g Schematic of a multilayer RF/d.c. spintronic neural network. The input RF signal is multiplied by the weight of individual MTJ synapses to generate d.c. voltages. The d.c. voltages will add up and be injected to the MTJ neurons so that RF signals can be generated and transmitted to the next layer of neural network. Parts (b, c) reprinted with permission from ref. ¹⁸, Springer Nature Limited. Parts (d–g) reprinted with permission from ref. ²¹¹, Springer Nature Limited.