A nested U-shaped network for accurately predicting directional scattering of all-dielectric nanostructures

Forward prediction of directional scattering from all-dielectric nanostructures by a two-level nested U-shaped convolutional neural network (U2-Net) is investigated. Compared with the traditional U-Net method, the U2-Net model with lower model height outperforms for the case of a smaller image size. For the input image size of 40 × 40, the prediction performance of the U2-Net model with the height of three is enhanced by almost an order of magnitude, which can be attributed to the more excellent capacity in extracting richer multi-scale features. Since it is the common problem in nanophotonics that the model height is limited by the smaller image size, our findings can promote the nested U-shaped network as a powerful tool applied to various tasks concerning nanostructures.