Probabilistic Representation and Inverse Design of Metamaterials Based on a Deep Generative Model with Semi‐Supervised Learning Strategy

Abstract

mostly relies on physics-inspired methods, resorting to human knowledge such as physical insights revealed by simplified analytical modeling, similar experience transferred from previous practice, and intuition obtained by scientific reasoning. For example, many meta-atoms inherited traditional antenna designs with geometries like rectangle, [4] cross, [5] bowtie, [6] V-shape, [7] H-shape, [8] and so on, whose first-order response is approximated by electrical dipole resonance with relevant scaling effect. [9] Some other designs guided by physical intuition include ring-like structures that exhibit strong magnetic resonances induced by the incident magnetic field, [10][11][12] dielectric building blocks that can induce both electric and magnetic resonances leading to better control of the phase of the scattered light, [13][14][15] or the spectra line-shape tailoring by introducing coupling among different resonant modes. [16][17][18] Despite the exciting results obtained by these physics-inspired designs, this methodology basically relies on a trial-and-error process, usually involving numerical methods like finite-difference-time-domain (FDTD) or finite element method (FEM) to iteratively solve Maxwell's equations. The low efficiency and thus limited exploration of the design varieties tend to easily omit the optimal solution. The inverse design approaches start from the opposite end, and try to optimize certain objective functions describing the desired performance. [19,20] Common approaches for inverse problems include genetic algorithm, [21] level set methods, [22] and topology optimization, [23] which, however, are still stochastic searching algorithms that are time-consuming and deteriorate rapidly as the design space grows. Different from numerical calculations, data-driven methods based on machine learning (ML) solve the optimization problem from statistical perspectives, so that the solution to optimize a target can be approximately generalized from numerous design examples. With the rapidly accumulated data and thus booming of deep learning (DL), the state-of-theart in many research domains, such as speech recognition, [24] computer vision, [25,26] natural language processing, [27] and decision making, [28] has been pushed far beyond conventional methods. Deep neural networks simulate biological signal processing that allow computational models to learn multiple The research of metamaterials has achieved enormous success in the manipulation of light in a prescribed manner using delicately designed subwavelength structures, so-called meta-atoms. Even though modern numerical methods allow for the accurate calculation of the optical response of complex structures, the inverse design of metamaterials, which aims to retrieve the optimal structure according to given requirements, is still a challenging task owing to the nonintuitive and nonunique relationship between physical structures and optical responses. To better unveil this implicit relationship and thus facilitate metamaterial designs, it is proposed ...

Design for quality: reconfigurable flat optics based on active metasurfaces

“…Due to the immense DOFs in constructing metasurfaces, early demonstrations using traditional NNs are restricted by the initial training sets with limited geometrical parameters. Generative adversarial network (GAN) methods, which are based on a self-evolving critic evaluating generated designs, allow efficient creation of arbitrary design patterns [319][320][321][322][323]. Such techniques customized for specific electromagnetic processes with complex metaatom geometries have also been developed [324][325][326][327][328].…”

Generative Deep Learning Model for a Multi-level Nano-Optic Broadband Power Splitter

“…A convolutional neural network (CNN) is shown to be effective in handling the geometrical input data 20,21,23,27 . Here we use two convolutional layers and the channel for them are 16 and 32, and the max pulling is 2, after that there is one fully connected layer to reduce the latent variable to 63.…”

“…Taking inspiration from these algorithms, machine learning (including deep learning) is becoming more popular to assist the design process of photonic devices to accelerate the underlying optimization process in the past few years. Recent success of deep learning in modeling complex input-output relationship in spatial-temporal data, has inspired the idea of intuitive physics engines that can learn physical dynamics in mechanics [7][8][9] , material discovery 10-12 , particle physics 13 , and optics [14][15][16][17][18][19][20][21][22][23][24][25] .…”

“…Liu et al has applied the Generative Adversarial Networks (GAN) 15 . and Ma et al has employed the Variational Autoencoder (VAE) 23 for their applications. Inspired by these works, we propose to utilize Conditional Variational Autoencoder (CVAE) in our power splitter application.…”

Prediction Network of Metamaterial with Split Ring Resonator Based on Deep Learning

“…At the same time, successes in other non-computer related fields are numerous, including many basic disciplines such as life sciences, chemistry [19], and physics [20] [21]. Therefore, applying deep learning to the design of metamaterials is also a hot research direction at present, and many outstanding works have appeared [22][23][24].…”