Prediction of optical properties of uniaxial hyperbolic nanospheres via artificial neural network

Abstract In this study, absorption and scattering multilayer perceptron models are developed and validated to predict the optical spectra of uniaxial hyperbolic nanospheres. The finite difference time domain method is used to generate the dataset of the absorption and scattering optical spectra. The models’ optimal performance is achieved for 5 hidden layers and 80 neurons for the absorption and the same hidden layers with 120 neurons for the scattering. The predictions by the model on the test dataset give a low average mean squared error of 0.000145 for the absorption and 0.00071 for the scattering. We also performed a robustness test by using parameters outside the initial parameters used for the training and the predictions are in good agreement with the actual datasets for both absorption and scattering. This research shows the application of artificial neural networks to predict the optical properties of hyperbolic materials and lays the groundwork for developing more complicated neural network models to predict complex phenomena in hyperbolic metamaterials, which are faster and computationally less expensive than using conventional simulation methods. Hyperbolic metamaterials offer unique optical properties that can be used to design new optical devices.