Deep neural networks for inverse design of multimode integrated gratings with simultaneous amplitude and phase control

Abstract

We present a photonic mode converter based on a grating structure, modeled and inversely designed by deep neural networks. The neural network maps the physical parameters of the grating to the grating responses, i.e., complex scattering parameters representing the reflected modes from the grating structure. We design different neural networks to output the magnitudes and the phases of the scattering parameters associated with the multiple reflected modes. Following the training process, we use the trained networks to perform inverse design of the grating based on the desired magnitudes of the scattering parameters. The inverse design effort provides a full control on the magnitudes and the phases of the reflected modes from the mode converter. Our techniques help in creating a rich landscape of multiple interfering waves that provide higher control on optical near fields, complex resonators, and their relevant nanophotonic applications.