Study On The Silicon-Based Nonlinear Activator For Photonic Neural Networks

Due to the development of Artifcial Intelligence(AI),the electronic signal processing hardware used are required to improve their performances.Compared with traditional electronic hardware accelerators,photonic hardware accelerators have the advantages of low power consumption,low latency,large bandwidth and high parallelism in optical transmission.Meanwhile,siliconbased photonic devices have the advantages of high integration,CMOS compatibility.Therefore,silicon-based photonic neural networks have become one of the developing trends.The siliconbased photonic neural networks utilize photons as the physical media to build the basic computing units of the AI computing,including linear matrix-vector multiplications and nonlinear activation functions.At present,the research mainly focuses on the linear part,while the nonlinear activation is mostly realized by photoelectric conversion to complete in computer.However,the photoelectric conversion process will lead to additional delay,loss,and increase the complexity of the peripheral circuit.Thus,the realization of optical nonlinear activators has gradually become one of the hot research directions.In this thesis,the research on the silicon-based nonlinear activators for optical neural networks is conducted,utilizing the strong nonlinearity of silicon and the high integration of silicon-based photonic devices.Through simulations and experiments,the performance of nonlinear activators based on waveguides and all-pass microring resonator(MRR)is verifed.Besides,the tunable nonlinear activator based on Fano spectrum of double beam MRR is designed and simulated.The main contents of this thesis are as follows:Firstly,two diferent simulation methods are proposed to verify the nonlinear characteristics of silicon-based waveguides and all-pass MRR respectively,considering the two situations of silicon-based ultrafast third-order nonlinearity dominated and thermo-optical efect dominated,which providing a theoretical basis for the realization of silicon-based nonlinear activators.When the thermo-optical efect can be ignored,the fnite diference time domain method is proposed to model the third-order nonlinearity of silicon using the Lumerical FDTD software,and the siliconbased waveguide is simulated.The saturation characteristics,negative diferential phenomena,spectrum splitting and other nonlinear phenomena are analyzed by simulating and testing.Since MRR has the feld enhancement efect,the thermo-optical efect cannot be ignored,so the timedomain coupled mode theory(TCMT)is used to simulate the all-pass MRR,and verifed through experiments.Secondly,the nonlinear activators based on waveguide and all-pass MRR are proposed for the diferent nonlinear functions extracted through the devices,realizing a low-power,cascadable all-optical nonlinear activator.The silicon-based waveguide is used as a nonlinear activator to extract the sigmoid function,whose recognition accuracies used for MNIST’s handwritten digit recognition is as high as 94 %.To further reduce power consumption,the nonlinear activator based on all-pass MRR is proposed,which can extract three kinds of nonlinear functions: GELU,Radial bias,and peak function.All the recognition accuracies of MNIST handwritten numerals are higher than 90 %.Besides,the power threshold can be reduced to 0.602 m W.Further,the TCMT is used for transient simulation to analyze the efect of self-induction oscillation on response time,and verifed through experiments.The cascading property of nonlinear activators is also discussed.By increasing the power to 3m W and the radius to 30 μm,the number of cascaded layers can be increased to 8.Thirdly,to meet the requirements of multiple functions and fxed input wavelengths of the neural network,it is proposed that the Fano spectrum and thermal tuning are combined to extract four nonlinear functions in single port and fxed wavelength : Sigmaid,GELU,Radial bias and peak function.The recognition accuracies of MNIST handwritten digits are around 90 %,providing a design idea for the tunable nonlinear activator.