AI-Based Fast Design for General Fiber-to-Waveguide Grating Couplers

Utilizing deep learning to replace numerical simulation solvers for electromagnetic wave propagation is a promising approach for the rapid design of photonic devices. However, to realize the advantages of deep learning for rapid design, it is essential to apply it to a general device structure. In this study, we propose a method that employs deep learning to assist in fast design of a general grating coupler structure. We use a modified 1D-ResNet18(1D-MR18) to predict the coupling efficiency of various grating couplers at different wavelengths. After comparing and selecting the optimal combination of learning rate, activation functions, and batch normalization size, the 1D-MR18 demonstrates remarkable accuracy (MSE: 2.18 × 10−5, R2: 0.969, MAE: 0.003). By integrating the 1D-MR18 with the adaptive particle swarm algorithm, we can efficiently design periodic and nonuniform grating couplers that meet various functional requirements, including single-wavelength grating couplers, multi-wavelength grating couplers, and robust grating couplers. The time for designing a single device is no more than 2 minutes, and the shortest is only 17 seconds. This novel approach of employing deep learning for the fast and efficient design from standard photonic device structures offers valuable insights and guidance for photonic devices design. © 2023, Electromagnetics Academy. All rights reserved.