| In the digital economy era,the total volume of global data and the scale of arithmetic power have exhibited explosive growth trends,escalating the demands on processor capabilities.As transistor miniaturization approaches its physical limits and Moore’s law slows,current computing paradigms encounter challenges,including insufficient computational power and high energy consumption.Benefiting from the inherent parallelism and interconnectivity advantages of optics,analog optical computing,exemplified by optical neural network（ONN）,presents a feasible solution to overcome the technological bottlenecks of the post-Moore era.ONN combines neural networks with optical information processing,centering on mapping computational processes onto an appropriate optical hardware platform.Physically,each layer of the ONN consists of an optical interference unit（OIU）that realizes matrix multiplication and an optical nonlinearity unit（ONU）that implements nonlinear mapping.Given that linear operations account for most of the computational cost in neural networks,most proposed ONN solutions adopt optoelectronic hybrid architectures,where the OIU is implemented optically while the nonlinear activation is done in the electrical domain.Although this model alleviates the computational burden on electrical processors,the process involves optical-to-electrical（O/E）conversion,leading to increased energy consumption and a lower computing rate,thereby failing to fully utilize the advantages of optical computing.On the other hand,unlike optical matrix calculations,implementing nonlinear operations in the optical domain remains a significant challenge.Therefore,this work focuses on the optical realization of nonlinearity and explores all-optical nonlinear activators for ONN.The main research of this dissertation is as follows:（1）To address the O/E rate mismatch and additional energy consumption in existing hybrid computing architectures,an all-optical nonlinear activator based on molybdenum disulfide（Mo S2）is designed.Profiting from the ultrafast high third-order nonlinear susceptibility and carrier dynamics of Mo S2,the nonlinear activation model based on saturable absorption is experimentally structured.For comparison,the absorption characteristics of Mo S2 films with different thicknesses are measured.On this basis,a rough mapping model between material thickness and nonlinear coefficient is also built.Meanwhile,a simulation-based fully connected neural network is fabricated to mimic the operation of ONNs and illustrate the feasibility of Mo S2 as an ONU.The results show that,for the handwritten digit classification task,the recognition accurateness ranged from 89%to 94%,depending on the morphology of Mo S2.（2）To solve the service life and weak nonlinearity limitations of the Mo S2 activator,an all-optical nonlinear activator based on a direct-coupled fiber ring resonator（DCFRR）is proposed.DCFRR,composed of simple passive components,enhances the nonlinear effects of input and output light intensities using a ring resonant structure.On this basis,the static and dynamic transmission features of the DCFRR activator are experimentally analyzed.The device demonstrates excellent nonlinear properties,high stability,versatility,and compatibility with existing optical chips,which benefits the scalability of ONNs.As a proof of concept,the performance of FORR as a nonlinear unit in ONNs is demonstrated.For image classification and super-resolution reconstruction tasks,the experimental performance of the model is comparable to commonly used activation functions in computers.（3）To obtain stronger nonlinear mapping,an all-optical nonlinear activator based on stimulated Brillouin scattering（SBS）is demonstrated.Considering the Brillouin scattering effect in fiber media,the SBS activator exhibits two different nonlinear mapping models in the forward transmission and backward scattering directions,corresponding to saturable absorption and Re LU models,respectively.Experimental verification shows that the proposed activator has a large dynamic response bandwidth（～11.24 GHz）,low threshold（～2.29 m W）,high stability,and wavelength division multiplexing characteristics.These features have potential advantages for the physical realization of optical nonlinearities.Likewise,we verify the performance of the proposed activator as an ONN nonlinear mapping unit via numerical simulations.Simulation shows that,under the role of hard-parameter sharing,our approach implements comparative advantages over the classical activation functions in the image classification and similarity comparison tasks,guiding more flexible applications of neural networks.This dissertation explores the interaction mechanisms between light and matter,verifies the physical realization of optical nonlinearity,and provides novel insights into nonlinear processing in the optical domain.On this basis,the feasibility of the nonlinear activator within ONN architecture is further demonstrated,which provides support for overcoming the limitations in the existing optoelectronic hybrid computing model and fully leveraging the advantages of optical computing.The related research content provides new solutions for exploring efficient,reliable and practical all-optical nonlinear activators. |
