Nanoplasmonic waveguide design driven by multilayer perceptron artificial neural network algorithm

Deep learning has gained growing popularity in multiple research fields by utilizing algorithms and statistical models to reveal underlying patterns of data that are collected from a problem, as distinct from its tedious mathematical calculation-related or time-consuming simulation counterparts. In this work, we apply a data-driven approach to the analyses of two nanoplasmonic waveguides, namely the Conductor-Gap-Dielectric (CGD) model and the Conductor-Gap-Conductor-Dielectric (CGCD) model. We first collect data from these two models via COMSOL and preprocess the acquired data accordingly. Then multilayer perceptron (MLP), which is the core of our data-driven approach, is used on the CGD model to determine the appropriate parameter settings of the machine learning model which leads to the combination of best-performing parameters that later on is applied to the analyses of CGCD model. The absolute percentage error of our algorithm is less than 4% for moderate parameter settings and could reach less than 1% when it is optimal. The algorithm also expressed great consistency with its parameters. From this work, it is seen that deep learning supersedes pure numerical simulations of a nanophotonic waveguide, especially in time efficiency. Machine learning has great potential in achieving faster and more accurate waveguide design.