Evaluation of Deep Neural Networks for Predicting Optical Properties of Silicon-rich Silicon Nitride Waveguide

Abstract

Deep learning (DL) has recently emerged as a potential platform for estimating linear and nonlinear optical phenomena of waveguides due to its high computational power, high-level structures and flexible usages. In this work, we performed a comparative analysis of four DL based Deep Neural Network (DNN) configurations for predicting and analyzing the effective mode area of a planar Silicon-rich Silicon Nitride (SRN) waveguide, its nonlinear coefficient, effective index and dispersion in the wavelength range of 0.65 µm – 3.05 µm, waveguide core width of 1 µm – 5 µm and waveguide height of 0.3 µm –0.4 µm. We found that out of four DNN structures analyzed, ELU-ELU-ReLU-70-9000 structure showed superior performance in terms of mean squared error values. The computational time required with deep neural network (for training) and finite-element method (FEM) solutions is also compared and found that the training time of DNN structures increased with a number of epochs and due to the ReLU activation function. This simple and fast-training DNN employed here predict the output for unfamiliar parameter setting of the optical waveguide faster than traditional numerical simulation techniques.