Machine learning-driven prediction of optical responses and inverse design of annular aperture arrays

Abstract

Annular aperture arrays (AAAs) in metallic films exhibit unique optical properties, making them promising for various nanophotonic applications. However, their design and optimization are computationally intensive due to the complex geometries and multi-parameter dependencies. In this paper, we employ artificial neural networks (ANNs) to predict the transmission spectra of AAAs. By training the ANN on a data set consisting of AAA geometric parameters and their corresponding transmission spectra, we demonstrate the ability of the model to accurately predict the optical response of AAAs with a validation mean squared error of just 0.0009. We investigate the underlying physical mechanisms responsible for the observed transmission peaks and their dependence on the periodicity of the AAA, confirming the findings of previous experimental and simulation studies. We further validate the model by comparing its predictions with the experimental data from AAAs used as color filters, highlighting its ability to capture key resonance peak position features. In addition, we introduce a dense geometry search strategy for inverse design, which enables rapid identification of optimal structures and achieves over 6000 times speedup compared to brute-force simulation. This work highlights the potential of deep learning in accelerating the design and optimization of complex nanophotonic structures, paving the way for accelerated development of next generation nanophotonic devices.