TCO‐Based Active Dielectric Metasurfaces Design by Conditional Generative Adversarial Networks. Issue 2 (27th November 2020)

• Record Type: Journal Article
• Authors: Jafar‐Zanjani, Samad; Salary, Mohammad Mahdi; Huynh, Dat; Elhamifar, Ehsan; Mosallaei, Hossein
• Is Part Of: Advanced theory and simulations. Volume 4:Issue 2(2021)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2020-11-27
• Subjects: active metasurfaces -- conditional generative adversarial networks -- indium tin oxide -- machine learning -- nanophotonic inverse‐design -- transparent conducting oxides; Science -- Simulation methods -- Periodicals; Science -- Methodology -- Periodicals; Engineering -- Simulation methods -- Periodicals; Engineering -- Methodology -- Periodicals; 507.21
• DOI: 10.1002/adts.202000196
• Languages: English
• ISSNs: 2513-0390

Abstract

Abstract: While researchers in the field of active flat optics continue to make groundbreaking progress by seeking novel materials and control systems, the complexity and sensitivity of the nanostructures that they aspire to design are unavoidably increasing. Inverse design of the popular class of transparent conducting oxide (TCO)‐based active metasurfaces is particularly challenging, largely due to the limited choice of the active materials, and sensitive physical mechanisms that give rise to their tunability. In this contribution, a new machine learning method based on a combination of the K‐means clustering algorithm and conditional Wasserstein generative adversarial networks (cWGANs) for broadband multi‐modal inverse design of TCO‐based active metasurfaces is developed. Multi‐objective evolutionary optimization is adopted to efficiently generate a diverse training dataset of high‐performance active metasurfaces. The training dataset includes samples that operate at specific wavelengths throughout the optical telecommunications (telecom) band. K‐means algorithm is then used to extract the clusters (modes) present in the training dataset, and exclusive cWGAN models are fit on each of them. The model is capable of generating designs operating at wavelengths that are not present in the training dataset. It also provides a clear picture of the feasibility and interplay between the design objectives. Abstract : A machine learning strategy for broadband multi‐modal inverse design of TCO‐based active metasurfaces is developed. Multi‐objective GA along with RCWA is used for generating a dataset of high‐performance designs at specific training wavelengths throughout the telecom wavelengths. K‐means algorithm is then employed to extract the clusters present in the dataset and wavelength‐conditioned cWGAN‐GP models are fit on the separated clusters.