Deep Learning-Based Modelling of Complex Photonic Crystal Slow Light Waveguides

Photonic crystal waveguides (PCWs) provide an effective way to manipulate light at sub-wavelength scale, but suffer from design difficulties such as long simulation time, low efficiency, and optimization challenges in the search for optimal values. We proposed an ultrafast solution based on the forward and reverse modeling of PCWs via a tandem classification regression neural network, that effectively solved the trouble of class imbalance in data and realized instant prediction of PCW's performances. The network has high accuracy as the coefficient of determination reaches 0.99. Furthermore, the reverse-designed neural network enables the determination of PCW parameters according to the desired optical properties. For instance, a PCW with a normalized delay bandwidth product of 0.458 and a group index of 71 was obtained. Our results could pave the way for future deep learning-based photonic device design towards high-performances.