A Generative Meta‐Atom Model for Metasurface‐Based Absorber Designs

Abstract

dielectric loss, [5] ohmic loss, [13][14][15][16] loading tunable lumped elements, [17][18][19][20] or multilayered metasurface. [15] Furthermore, polarization conversion and interference cancellation are introduced to the designs to achieve even better absorption and wider bandwidth. [21][22][23] Though the metasurface increases the design freedom of absorbers, it also causes complexity in design. Firstly, the increased dimensionality of the design freedom rapidly raises the cost of optimizations since the acquisition of EM response of MSA usually involves complex and time-consuming numerical simulations. [24] Next, the pattern of meta-atoms is usually given by physics-inspired considerations, which leads to regular patterns or their geometric transformations, such as split rings, [25] patches, [26] cross, [27] V-shapes, [28,29] and so on. These regular patterns are in a very small subspace of whole patterns, and thus the absorbers implemented by such patterns may not achieve the best absorbing performances.To meet the complexity caused by meta-atoms, one trend is using the deep learning (DL) method to model the meta-atoms, which have been extensively studied in computer science and engineering. [30] The unique feature of DL is that it can discover useful information from large amounts of data and establish data-driven models for the design. [24] For example, deep neural network (DNN) can capture data features with higher levels of abstraction in lower-level features and fit complex input-output relationships. Using this capability of the DL, one can construct a model to characterize the complex relationship between geometric topology and EM response for MS, and then timeconsuming numerical modeling is avoided in the design process. [31] This surrogate model will greatly improve design efficiency. However, it can only apply to the optimal design of MS with given topologies because of the one-to-one correspondence between the topologies and topological parameters. To overcome this shortcoming, the generative networks model attracts much attention. The generative model can not only effectively capture the essential features of complex input datasets, but uses the feature knowledge to generate new data. For example, researchers build generative adversarial networks (GANs) to implement the inverse design of metasurface patterns. [32][33][34] However, the inverse design model meets the convergence difficulty because of the existence of multiple candidate solutions.In this work, we propose a generative meta-atom model for the generation and optimization of MS toward absorbing Metasurfaces (MS) are widely accepted in the devices, such as absorbers, to improve their performances. MS provide new design freedoms, however, they also result in complexity in designs because of the high-dimensional topological space of meta-atoms, and high computational costs in the optimization procedure. To alleviate this challenge, a generative meta-atom model that generates the pattern and corresponding electromagnetic (EM) responses of m...

Designing Metasurfaces for Efficient Solar Energy Conversion

“…When it comes to the generative networks application in desired metasurface designs, Ding et al established a model by the combination of convolutional variational autoencoder and deep neural network to realize ultrabroadband, low-profile absorbers design . In addition, generative networks have been combined with topology-optimized designs for the rapid generation of highly efficient metasurface designs.…”

“…179 So et al also reported a deep learning method, which facilitates highly robust spectrally sensitive multiband absorbers that has a low average mean squared error (MSE, Figure 6d). 180 When it comes to the generative networks application in desired metasurface designs, Ding et al established a model by the combination of convolutional variational autoencoder and deep neural network 181 to realize ultrabroadband, low-profile absorbers design. 181 In addition, generative networks have been combined with topology-optimized designs for the rapid generation of highly efficient metasurface designs.…”

“…180 When it comes to the generative networks application in desired metasurface designs, Ding et al established a model by the combination of convolutional variational autoencoder and deep neural network 181 to realize ultrabroadband, low-profile absorbers design. 181 In addition, generative networks have been combined with topology-optimized designs for the rapid generation of highly efficient metasurface designs. Kudyshev et al introduced several methods that combine topology optimization with deep learning approaches for the inverse design of metasurface thermal emitter for TPV applications (Figure 6e−g).…”

Design of Multifunctional Color Routers with Kerker Switching Using Generative Adversarial Networks

“…Notably, image-based generative networks offer a high degree of design freedom and have played a significant role in optical design endeavors, such as unit cells in metasurfaces. [33][34][35][36][37][38][39][40] In this study, we present a novel approach based on generative adversarial networks (GANs) for designing single-element multifunctional color routers. Given the complexity of coupling between electromagnetic multipole modes, the spectrum of a dielectric resonator represents a complex combination of Fano and Lorentz lineshapes.…”

Deep-learning-assisted design of multi-degree-of-freedom metamaterial absorber

“…Typically, the unit structure is designed with a regular shape and a single material; therefore, designers must have extensive experience in this area. In recent years, deep learning has simplified and improved the efficiency of metamaterial device design, which is becoming a popular design methodology [23][24][25][26][27]. Even though the deep learning-based design method improves the design efficiency of MAs, it still suffers from a low degree of freedom in the design of unit structures.…”