Study On The Key Technology Of Silicon-Based Photonic Devices And Systems For Photonic Neural Networks

Due to the increasing accumulation of big data,various algorithms of artificial intelligence(AI)are coming out.The existing electronic signal processing hardware used for AI technology are required to improve their performances.Optical computing and photonic neural networks as the new methods are proposed to balance the demands between the substantial increase in computation ability and the decrease in computing power consumption,owing to its superiorities in terms of large bandwidth,low loss,low latency,low power consumption and high parallelism in optical transmission.Besides,the silicon-based material system has the advantages of CMOS compatibility,high integration,optoelectronic monolithic integration,and it can provide the mature fabrication technology for large-scale optical neural network integration circuit.Therefore,in this thesis,the systematic researches on the silicon-based photonic devices and its applications for photonic neural networks are conducted.The silicon-based 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.In order to realize the on-chip integration of the basic computing units and improve the flexibility of its large-scale expansion,microring resonators are used as the fundamental devices with compact footprint and low power consumption.In addition,the technology of wavelength division multiplexing(WDM)is also introduced.Based on the above-mentioned devices and technology,silicon-based optical linear matrix-vector multiplication and all-optical nonlinear activator are proposed and demonstrated.Furthermore,based on the realization of this architecture,the application of silicon-based incoherent optical neural network in real-time temperature sensing for the distributed optical fiber sensing system is also demonstrated,which might provide a feasible solution for the practical application of photonic neural networks.The main contents and contributions of this thesis are as follows:Firstly,a silicon-based Mach-Zehnder interferometer(MZI)-embedded microring resonator(MZER)is proposed and used as the fundamental device of the silicon-based incoherent optical linear matrix-vector multiplication.In one aspect,the proposed MZER can be utilized to realize the”hitless” weight configuration.It is worth noting that during the tuning process of adding/dropping a weighted channel in a WDM network,it would encounter a problem that the other weighted channels will be interrupted and the unexpected wavelengths will be dropped.The embedded MZI in the microring resonator works as an ON/OFF key of the resonance,and can allow hitless weighted wavelength switching without FSR-alignment.Moreover,it can offer a robust adjustment of the applied voltage power,owing to the large operating wavelength range provided by the equal-armed MZI.As for the four weighted channels,the transmission performances of four channels keep with good consistency within 35.2 nm.Besides,the crosstalk of the four channels is around 20 d B.In another aspect,the proposed MZER can realize the weight configuration with high precision,which is based on the larger extinction ratio introduced by the structure of the equal-armed MZI.The experimental results verify that the MZER structure can realize the weight configuration with simple operation,high accuracy and 6-bit precision.In further,a silicon-based MZI-embedded microring array consisting of four weighted channels are designed and fabricated.The accuracy and precision of each weighted channel can be maintained above 5.5 Bits,in which the input power is-4 d Bm.Besides,the accuracy and precision could be improved with the growth of the input power.Secondly,based on the numerical analysis model for extracting the nonlinear characteristics of the silicon-based devices,it is proposed that the silicon-based waveguides and the all-pass microring resonators can be used as silicon-based all-optical nonlinear activators with different nonlinear functions.The nonlinear performances of the silicon-based waveguides and microring resonators with different parameters are simulated and analyzed.It is verified that the correspondences between the input and output optical power satisfy different nonlinear functions.In further,these different nonlinear functions are applied to the fully-connected neural networks,and used for the recognition of MNIST handwritten digits.The recognition accuracies based on different silicon-based devices all can reach more than 90%.Thirdly,considering the characteristics of the silicon-based photonic neural network,new applications on optical sensing and the real-time signal processing are investigated.In distributed optical fiber sensing,the real-time temperature sensing of Brillouin Optical Time Domain Reflectometry(BOTDR)is one of the most widely used applications.In order to avoid the problems brought about by the traditional algorithms including long processing time,insufficient accuracy and robustness,a wavelet convolutional neural network(WNN)for robust and fast temperature measurement in BOTDR is proposed and demonstrated.Both the simulated and measured results show that the WNN has better robustness and flexibility.Besides,the computational accuracy of the WNN is improved and the fluctuation of that is slighter,especially when the SNR is less than11 d B.In addition,the proposed WNN can be improved into three small-scale cascaded siliconbased photonic neural networks,which only consist of linear matrix-vector multiplication and nonlinear activation.Hence,the temperature sensing could be realized by the improved cascaded silicon-based photonic neural networks.As a result,its root mean square error and standard deviation of extracted temperature could be maintained in the range of 1.5℃-2℃,which meets the performance requirements in the distributed optical fiber sensing system.