Modeling and Characterization of All-Dielectric Metasurfaces with High Degrees of Freedom Based on Deep Residual Network

The principal challenge in metasurface design is the non-intuitive design where the response properties of meta-atoms dictate the design of their structure. In this paper, we present a high-degree-of-freedom all-dielectric metasurface modeling approach based on a deep residual neural network that includes design parameters such as free-form structure, dielectric thickness, refractive index and lattice size. We designed an efficient network architecture to predict the real part and imaginary part of the transmission coefficients at 301 sampling points within tens of milliseconds, yielding the transmission amplitude and phase for 30–60 THz. In generalization performance test, the mean square errors of real and imaginary part prediction are 0.00092 and 0.00094 respectively, while the mean square error of the amplitude prediction is 0.00081 and the mean absolute error of the phase prediction is
$2.52^{\circ}$
. The prediction network accelerates the prediction process and achieves high accuracy, making it a highly practical tool.