在本研究中，我們在晶片上設計了一個全光學的類神經網路架構，用來實現基於延遲的reservoir computing (RC)，並從理論上研究其性能表現。此RC系統是被設計成製作在 Silicon-on-Insulator (SOI)晶圓上。依據IMEC-SiPh (iSiPP50G) 矽光子晶片製程規範而設計。晶片上的奈米光學元件包含：耦合器、波導、螺旋波導、馬赫-曾德爾調變器、相位調變器和光柵耦合器。 本研究利用耦合器中光的干涉提供在RC中所需的非線性函數效果。而奈米光學元件和RC系統的性能表現將被評估和優化。 方波和三角波光訊號會被送到此RC系統。當輸入訊號為方波時，RC系統的輸出訊號強度為高準位；當輸入訊號為三角波時，RC系統的輸出訊號強度為低準位。最低NRMSE為0.0864，由模擬結果可看出此RC系統具辨識輸入訊號的能力。;In this study, we design an on-chip all-optical neural networks architecture for a delay-based implementation of reservoir computing (RC) and investigate its performance theoretically. The RC system is designed to be fabricated on a Silicon-on-Insulator (SOI) wafer using the design rules of IMEC-SiPh (iSiPP50G). The nano-optics components used in the chip include directional couplers, waveguides, spiral waveguides, Mach-Zehnder modulator, phase shifters, and grating coupler. The nonlinear function required in RC is obtained by the interference of the light in the directional couplers. The performance of the nano-optics components and the RC system will be evaluated and optimized. The rectangular and triangular optical signals are launched into the RC system. The output intensity of the RC system is high and low level as the input signals are rectangular and triangular, respectively. The lowest value of NRMSE is 0.0864. The simulation results show that the RC system can recognize the input signals.
