Efficient inverse design and optimization of grating couplers using deep neural network models

Grating couplers enable efficient optical coupling between waveguides and optical fibers. Despite its critical impact on coupling performance, the influence of the coupling angle remains unclear in the existing study. We explore the influence of the coupling angle on the coupling efficiency spectrum of grating couplers from both theoretical and numerical simulation perspectives. Then, the rapid inverse design of grating couplers is accomplished using deep neural networks, which establish a nonlinear mapping between the coupling efficiency spectrum and key structural parameters. To ensure the inverse design consistency of the diverse optical devices, we introduce a comprehensive analysis of model hyperparameters to enhance the universality of the proposed method. Extensive experiments demonstrate that the proposed method can achieve a target coupling efficiency spectrum with the Prediction's Accuracy of up to 0.920 and MSE down to 0.0192. Notably, shifting the target spectrum results in a 24.6% and 30.4% increase in coupling efficiency at the central wavelength for the left-shifted and right-shifted cases, respectively, with coefficients of determination R2 of 0.997 and 0.998 between the shifted and the simulated coupling efficiency spectrum. We provide valuable insights and guidance for the future development of silicon photonic devices.