Machine Learning Approach for On-Demand Rapid Constructing Metasurface
Zhichao Sun, Bijun Xu, Fan Jin, Gongxue Zhou, Lin Lü
Abstract
Metasurfaces have developed rapidly with the extraordinary electromagnetic properties in electromagnetic wave control in recent years. However, the conventional metasurfaces design based on the Method of Moments (MOM), Finite Element Method (FEM) and Finite Integration Technique (FIT) are still time-consuming and demand significant computation. In this paper, we proposed a polynomial regression of standardized K-nearest neighbor algorithm (PS-KNN). The trained model shows an excellent prediction ability, the means square error (MSE) of the forward model is only 3.463 × 10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−6</sup> . We further report a reverse model based on forwarding prediction, which automatically constructs and optimizes the meta-atom by standardizing the electromagnetic properties (amplitude, phase, etc.) of the metasurface as the input of characteristic parameters. The MSE of the reverse model is 1.589 × 10 <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">−3</sup> . Finally, we cascade the two models, and predicted successfully eight meta-atoms by the closed-loop network and arrange them into a focused array. The results demonstrate the algorithm model avoids extensive modeling operations and numerical calculation and over 300 times faster than traditional electromagnetic simulation software. It offers a novel effective methodology to accelerate the on-demand design of complex metasurfaces and optical structures.