Deep learning based reconstruction of directional coupler geometry from electromagnetic near-field distribution

Abstract

We demonstrate a method to retrieve the geometry of physically inaccessible coupled waveguide systems based solely on the measured distribution of the optical intensity. Inspired by recent advancements in computer vision, and by leveraging the image-to-image translation capabilities of conditional generative adversarial neural networks (cGANs), our method successfully predicts the arbitrary geometry of waveguide systems with segments of varying widths. As a benchmark, we show that our neural network outperforms nearest neighbor regression from both a runtime and accuracy point of view.