Generative Deep Learning Model for Inverse Design of Integrated Nanophotonic Devices
Yingheng Tang, Keisuke Kojima, Toshiaki Koike–Akino, Ye Wang, Pengxiang Wu, Youye Xie, Mohammad H. Tahersima, Devesh K. Jha, Kieran Parsons, Minghao Qi
Abstract
Abstract A novel conditional variational autoencoder (CVAE) model for designing nanopatterned integrated photonic components is proposed. In particular, it is shown that prediction capability of the CVAE model can be significantly improved by adversarial censoring and active learning. Moreover, generation of nanopatterned power splitters with arbitrary splitting ratios and 550 nm broadband optical responses from 1250 to 1800 nm are demonstrated. Nanopatterned power splitters with footprints of 2.25 × 2.25 m 2 and 20 × 20 etch hole positions are the design space, with each hole position assuming a radius from a range of radii. Designed nanopatterned power splitters using methods presented herein demonstrate an overall transmission of about 90% across the operating bandwidth from 1250 to 1800 nm. To the best of authors' knowledge, this is the first time that a state‐of‐the‐art CVAE deep neural network model is successfully used to design a physical device.