Neural network modeling and uncertainty quantification of photonic crystals incorporating an anisotropic defect layer

Abstract This study presents a machine learning (ML) framework for rapid prediction and uncertainty quantification (UQ) of optical properties in complex photonic structures using a multilayer perceptron (MLP) neural network. We investigate a one-dimensional photonic crystal incorporating a graphene-hyperbolic metamaterial defect layer with a tilted optical axis, which induces unique anisotropic optical responses. A comprehensive dataset spanning a broad range of structural and material parameters was used to train the MLP model. Once trained, the MLP rapidly and accurately predicts transmission spectra and field distributions, capturing dual-peak defect modes within the bandgap and spatial field localization, thereby enabling efficient analysis of complex photonic structures without repeated numerical simulations. Rigorous UQ reveals outstanding predictive reliability with R 2 scores exceeding 0.97 and well-calibrated prediction intervals achieving 89.8% coverage probability. This work establishes ML as a powerful paradigm for the design and optimization of advanced photonic devices using hyperbolic metamaterials and anisotropic media, offering unprecedented computational efficiency with reliable UQ for real-time device optimization and inverse design applications.