Search

Metasurfaces is one of the research hotspots in the field of micro/nanometer technologies during the last decade. Whereas, it remains challenging to inverse design optics devices according to desired physical response. Here we demonstrate an attempt to design a metasurface meta-atoms with machine learning approach. We use the metasurface structure parameters, nanofin material, phase, and transmission as model data of the deep neural networks (DNN) and light gradient boosting machine (LighGBM), and the mapping relationship between geometric parameters and phase and transmission is established. The two models are highly accurate in forward design, achieving the best regression coefficient of 0.969, while for the inverse design the best regression coefficient is 0.918. Using the trained inverse design network, the structural parameters and nanofin composition of the metasurface meta-atoms can be obtained for a given phase and transmittance, and the two models show strong generalization ability. This method may facilitate the viability of complex metasurface design and it can be also potentially applied in optical zoom imaging, perfect absorbers, metasurface filters, and other nanophysics fields. • The highlights of this paper are mainly the following three points. • We demonstrate the application of LightGBM and DNN algorithms in the forward and inverse design of metasurface meta-atoms. • The transmittance and phase are predicted in the forward design. • The nanofins and structural parameter of the metasurface meta-atoms is design by inverse design. [ABSTRACT FROM AUTHOR]

• The innovative use of multi-headed neural networks in combination with tandem neural networks solves the non-uniqueness problem of Layered Photonic Structures. • A self-attentive mechanism is added to the front of the inverse network to improve the accuracy and problem-solving capability of the network-aided design. • The generalization capability of the forward network is improved using cross-validation, which results in a significant improvement in the overall network performance. Neural networks have proven to be an influential tool in assisting with the inverse design of nanophotonic structures. However, the issue of non-uniqueness poses a significant limitation to this approach, as disparate designs can produce nearly identical spectra. This problem can result in the neural network failing to converge or producing erroneous results. In this study, we propose a multi-headed tandem neural network (MTNN) approach to address this issue. This method enables the neural network to generate multiple sets of outputs and utilize tandem neural networks (TNNs), and self-attention mechanisms, among other techniques, to constrain the results, and let these multiple outputs be fitted separately to different results. This allows the neural network to converge without sacrificing the simplex solution in the face of multimodal solutions. We employ the MTNN approach to inverse engineer a multilayer photonic structure comprised of two sets of oxide films, and the multiple outputs provide numerous valuable solutions. Our approach presents an effective solution for the inverse design of photonic structures afflicted with non-uniqueness problems. [ABSTRACT FROM AUTHOR]