Exploring High-Performance Slow Light Grating Waveguides By Means of Deep Learning

The geometric parameters of periodic units, which influence the refractive/reflective behaviors of grating waveguides, are usually designed using intuition combined with meticulous trial-and-error, making the design extraordinarily time-consuming and computationally intensive. In this work, we proposed an effective method based on deep learning to meet this challenge. Concretely, we employed fishbone slow-light waveguides as the object to be optimized, whose unit cell has multiple degrees of freedom, and trained a deep neural network (DNN) to predict their geometric parameters with the given discrete group index curves as the input. Our design is oriented for high group index and near-zero group velocity dispersion (GVD). To verify this functionality, we performed the controllable inverse design of GVD while keeping the group index at the same level (about 24), which made the bandwidth increase by 53.8% at the same time, and further explored near-zero GVD structures with a group index of 30.76/41.30/51.90. Our results lay the groundwork for the rapid design of complex slow-light grating waveguides.