Design and analysis of guided modes in photonic waveguides using optical neural network

TL;DR

This paper introduces an optical neural network method to efficiently predict fundamental modal indices in silicon waveguides, significantly reducing computational effort while maintaining high accuracy, and compatible with current fabrication technologies.

Contribution

The paper presents a novel deep learning approach using optical neural networks to predict waveguide modes with minimal simulations, enhancing efficiency and applicability in photonics design.

Findings

Mean squared error of predictions <10^{-5}

Compatible with CMOS fabrication ranges

Reduces simulations from thousands to a few hundred

Abstract

We present a deep learning approach using an optical neural network to predict the fundamental modal indices $n_{\rm{eff}}$ in a silicon (Si) channel waveguide. We use three inputs, e.g., two geometric parameters and one material property, and predict the $n_{\rm{eff}}$ for the transverse electric and transverse magnetic polarizations. With the least number (i.e., $3^3$ or $4^3$) of exact mode solutions from Maxwell's equations, we can uncover the solutions which correspond to $10^3$ numerical simulations. Note that this consumes the lowest amount of computational resources. The mean squared errors of the exact and the predicted results are $<10^{-5}$. Moreover, our parameters' ranges are compatible with current photolithography and complementary metal-oxide-semiconductor (CMOS) fabrication technology. We also show the impacts of different transfer functions and neural network layouts on the model's performance. Our approach presents a unique advantage to uncover the guided modes in any photonic waveguides within the least possible numerical simulations.