Stacked neural networks for predicting scattering spectra of core-(multi)shell particles

Artificial neural networks (ANNs) have been used in various nano-optical applications, e.g. in the fast prediction of the optical scattering response from various structures [1] . Here, we report on an approach to use ANNs to calculate the scattering cross section of core-(multi)shell particles, similarly as reported in [2] . For the allowed materials of the core-shell particle we limit them to a fixed number of (five) materials labelled by integers 1 to 5, each with a corresponding refractive index. A significant improvement with respect to the existing state-of-the-art is achieved by training the neural network in a sequential shell-by-shell manner, i.e. for a core-shell particle with 2 shells we would employ three networks, one is dedicated to the core and there are two further networks dedicated to each shell. All the "shell" networks have as additional input the scattering spectrum predicted from preceding layers. During the training phase, this will be taken from multishell Mie calculations [3] , and during prediction phase, this input will be provided by the respective ANN of the previous shell.