计算机科学
电子工程
光子学
谐振器
电压
炸薯条
人工神经网络
干涉测量
传输(电信)
CMOS芯片
光学滤波器
卷积神经网络
控制器(灌溉)
光子集成电路
芯片上的系统
低功耗电子学
控制系统
非线性系统
深度学习
信号处理
工程类
电力传输
作者
Hongsheng Niu,Dongdong Lin,Shangqing Shi,Feng Zhang,Nuoyi Zhou,Rui Ma,Junhao Ni,Chen Guo,Enze Zhou,Yifei Chen,Jin Wang,Wei Cheng,Guohua Hu,Yiping Cui,Binfeng Yun
标识
DOI:10.1109/jlt.2025.3627565
摘要
Reconfigurable photonic chips have attracted considerable attention for their broad applicability across various scenarios. However, manual control of these chips is difficult and time-consuming due to the presence of multiple electrodes and the inherent thermal crosstalk between them. To tackle this issue, some feedback-control methods based on additional on-chip optical power taps and self-configuration using intelligent optimization algorithms have been proposed. Although these methods are beneficial, they are constrained by low efficiency, limited flexibility, and a tendency to converge to local optima. In this work, we proposed and experimentally validated a deep learning-driven control scheme for reconfigurable photonic chips. The trained deep learning framework can rapidly and accurately predict the required control voltages based on user-specified target responses. As a proof of concept, a convolutional neural network (CNN) was implemented to establish the nonlinear mapping between the transmission spectra of a reconfigurable microring resonator (MRR)-assisted asymmetric Mach-Zehnder interferometer (AMZI) chip and its required control voltages through end-to end learning. Utilizing the trained CNN, the necessary control voltages for different target transmission spectra of the chip can be accurately predicted within 100 milliseconds. The experimental results demonstrated that the trained CNN attained a mean absolute percentage error (MAPE) of less than 0.5% on the test dataset. Moreover, an error margin of under 2% was achieved when utilizing user-defined simulated spectra as the target response. The proposed scheme provides a highly efficient approach for the intelligent control of reconfigurable photonic chips.
科研通智能强力驱动
Strongly Powered by AbleSci AI