Robust inverse-design of scattering spectrum in core-shell structure using modified denoising autoencoder neural network

Abstract

Neural network-based inverse design of nanophotonic device network is computationally and time efficient, but in general suffers the problems of robustness and stability against variation of the input target electromagnetic response. The inverse design network is required to be robust against the input electromagnetic response and to be capable of approximating the given electromagnetic response, even under the circumstances that the exact target response may not exist. We introduce a modified denoising autoencoder network to ensure the robustness of neural network-based inverse design, which consists of (1) a pre-trained network as a substitute of numerical simulation and (2) an inverse design network. We further purposely train the network with certain random disturbances added to the training dataset generated by the pre-trained network. Consequently, our modified denoising autoencoder network is more robust and more accurate than the conventional fully connected neural network. The strength and flexibility of our proposed network is illustrated via three concrete examples of achieving the desired scattering spectra of layered spherical scatterers.