Generation of Beam Profiles from Chip-to-Free-Space Coupling using Deep Neural Network

Three-dimensional (3D) simulation of light propagating from silicon photonics (SiPh) gratings to free space using the finite-difference time-domain (FDTD) technique can be time-consuming, with simulation times reaching up to 12 hours. To address this issue, we propose a deep neural network (DNN) modeling technique as a rapid, alternative method for visualizing light propagation. In this work, a DNN model is developed to generate beam profiles in chip-to-free-space coupling of SiPh gratings. The DNN model was trained only with 10% of the beam profiles from the FDTD-simulated 3D electric field (E-field), and used to generate the remaining 90% of the beam profiles. The total training time for the DNN is 17 minutes, while the time taken for E-field prediction and beam profile generation is less than 10 minutes. Comparing the FDTD-simulated and DNN-generated beam profiles, the average percentage error (APE) for beam waists is 4.69%, and the APE for the maximum E-field is 8.80%. When varying the architecture of the DNN model, prediction accuracies of > 0.94 and > 0.87 are obtained for beam waists and maximum E-fields, respectively.