A deep learning approach for inverse design of the metasurface for dual-polarized waves

Abstract

Compared to the conventional metasurface design, machine learning-based methods have recently created an inspiring platform for an inverse realization of the metasurfaces. Here, we have used the Deep Neural Network (DNN) for generation of desired output unit cell structures in an ultra-wide working frequency band for both TE and TM polarized waves. To automatically generate metasurfaces in a wide range of working frequencies from 4 to 45 GHz, we deliberately design an 8 ring-shaped pattern in such a way that the unit-cells generated in the dataset can produce single or multiple notches in the desired working frequency band. Compared to the general approach, whereby the final metasurface structure may be formed by any randomly distributed "0" and "1", we propose here a restricted output structure. By restricting the output, the amount of calculations will be reduced and the learning speed will be increased. Moreover, we have shown that the accuracy of the network reaches 91%. Obtaining the final unit cell directly without any time-consuming optimization algorithms for both TE and TM polarized waves, and high average accuracy, promises an effective strategy for the metasurface design; thus, the designer is required only to focus on the design goal. I. INTRODUCTIONMetamaterials, composed of subwavelength periodic or nonperiodic geometric arrays, have attracted widespread attention due to their peculiar assets to modify the permittivity and permeability 1,2,[27][28][29] . Recently, many novel functionalities have been implemented by metamaterials and their 2D counterpart, metasurfaces, such as intelligent surfaces for communication 3,4 , real-time wavefront manipulation 5-8 , perfect absorption 9,10 , and machine learning metasurface design 11,12 .However, all of these works are founded on conventional approaches including trial-and-error methods, brute force optimization methods, and parameter sweep, which are timeconsuming processes. Therefore, to solve the above challenges and to find a fast, effective, and automated route for metasurface design, we have benefited from machine learning. Deep learning is an efficient approach for learning the relationship between input and target data from the samples of past experiences. To be more specific, deep learning as a specific branch of machine learning can infer the basic rules based on formerly specified data; then, for another given input, the designed network can make reasonable decisions. With the ever-increasing growth of machine learning and its potential applications to address some important problems such as signal processing 13 and through the wall imaging 14 , we are now witnessing the opening of machine learning in wave-interaction phenomena. Owing to its potential capacity to provide higher accuracy, less design time, and improve the efficiency of a design process, machine learning has been introduced in numerous electromagnetic phenomena, for instance, all-dielectric metasurfaces 15 , antenna design 16,17 , acoustic metamaterials 18,19 , and computational el...

Intelligent metaphotonics empowered by machine learning

“…DL also opens the route for the design of biology-inspired devices such as moth-eye structures [73] with the designed average absorption reaching 90% in the range from 400 nm to 1600 nm. Scattering properties are a subject of many design procedures [56,[74][75][76][77][78], including those devoted to the development of anisotropic [79] and bianisotropic [80] metasurfaces, as well as switchable reflectors [59]. DL was also exploited for achieving electromagnetically-induced transparency [81][82][83].…”

Equivalent Circuit Theory-Assisted Deep Learning for Accelerated Generative Design of Metasurfaces

“…Apart from the forward design methods, many inverse design approaches [20][21][22][23][24][25][26][27][28][29] have been proposed and investigated. Contrary to the forward designs, EM responses of metasurface structures are set as the input while their corresponding geometrical parameters are the output of the neural network.…”

“…Once the inverse DNN is trained, geometrical parameters can be obtained directly given the target EM responses. Various metasurface designs, such as single-layer metasurfaces [20], [21], multi-layer metasurfaces [22], dual-polarized metasurfaces [23] and coding programmable metasurfaces [24], [25], have been reported demonstrating the feasibility of the inverse design methods. Compared to the forward design methods, the inverse neural networks provide a higher design efficiency.…”