Modal classification in optical waveguides using deep learning

Abstract

Single-mode operation is crucial in many on-chip integrated photonic devices, and thus the identification of single-mode geometries is an inevitable design requirement. In this article, we develop deep learning (DL) models for ultra-quick classifications of optical waveguide geometries into single- and multi-modal geometries. The DL model accurately predicts the boundary in the parameter space for the geometry of the waveguide that splits the space into single- and multi-modal regions. Using silicon nitride channel waveguide, and targeting both visible and telecommunication wavelengths, we illustrate how DL models can be developed with a minimal number of exact numerical simulations to Maxwell’s equations.