Generative Modelling of Three-Dimensional Beam Profiles from Grating Structure using Neural Network

In this study, a modelling technique to generate 3D optical beam profiles of chip-to-free-space coupling from silicon photonics (SiPh) gratings is demonstrated. The deep neural network (DNN) model training, electric field (E-field) prediction, and beam profiles generation of optimized DNN took ~1 hour, which is significantly shorter than the computational time of finite-difference time-domain (FDTD) that could take up to 12 hours. The average percentage error (APE) between FDTD-simulated and DNN-generated beam profiles reaches below 20% for beam profiles located at > 30 µm above the grating. This translates to >80% accuracy for DNN-generated profiles. Note that the source of E-field prediction error from DNN-generated profiles is mainly from the fringe-like irregularities in the FDTD-simulated profiles, which has minimal impairment when designing SiPh gratings for most chip-to-free-space applications.