Toward Machine-Learning-Accelerated Design of All-Dielectric Magnetophotonic Nanostructures

Abstract

All-dielectric magnetophotonic nanostructures are promising for integrated nanophotonic devices with high resolution and sensitivity, but their design requires computationally demanding electromagnetic simulations evaluated through trial and error. In this paper, we propose a machine-learning approach to accelerate the design of these nanostructures. Using a data set of 12 170 samples containing four geometric parameters of the nanostructure and the incidence wavelength, trained neural network and polynomial regression algorithms were capable of predicting the amplitude of the transverse magneto-optical Kerr effect (TMOKE) within a time frame of 10-3 s and mean square error below 4.2%. With this approach, one can readily identify nanostructures suitable for sensing at ultralow analyte concentrations in aqueous solutions. As a proof of principle, we used the machine-learning models to determine the sensitivity (S = |Δθres/Δna|) of a nanophotonic grating, which is competitive with state-of-the-art systems and exhibits a figure of merit of 672 RIU-1. Furthermore, researchers can use the predictions of TMOKE peaks generated by the algorithms to assess the suitability for experimental setups, adding a layer of utility to the machine-learning methodology.