Graphic-processable deep neural network for the efficient prediction of 2D diffractive chiral metamaterials.

Abstract

We propose a novel, to the best of our knowledge, graphic-processable deep neural network (DNN) to automatically predict and elucidate the optical chirality of two-dimensional (2D) diffractive chiral metamaterials. Four classes of 2D chiral metamaterials are studied here, with material components changing among Au, Ag, Al, and Cu. The graphic-processable DNN algorithm can not only handle arbitrary 2D images representing any metamaterials that may even go beyond human intuition, but also capture the influence of other parameters such as thickness and material composition, which are rarely explored in the field of metamaterials, laying the groundwork for future research into more complicated nanostructures and nonlinear optical devices. Notably, the rigorous coupled wave analysis (RCWA) algorithm is first deployed to calculate circular dichroism (CD) in the higher-order diffraction beams and simultaneously promote the training of DNN. For the first time we creatively encode the material component and thickness of the metamaterials into the color images serving as input of the graphic-processable DNN, in addition to arbitrary graphical parameters. Especially, the smallest intensity is found in the third-order diffraction beams of E-like metamaterials, whose CD response turns out to be the largest. A comprehensive study is conducted to capture the influence of shape, unit period, thickness, and material component of arrays on chiroptical response. As expected, a satisfied precision and an accelerated computing speed that is 4 orders of magnitude quicker than RCWA are both achieved using DNN. This work belongs to one of the first attempts to thoroughly examine the generalization ability of the graphic-processable DNN for the study of arbitrary-shaped nanostructures and hypersensitive nanodevices.