A Deep Learning Approach for Objective-Driven All-Dielectric Metasurface Design

Abstract

Metasurfaces have become a promising means for manipulating optical wavefronts in flat and high-performance optical devices. Conventionally metasurface device design relies on trial-anderror methods to obtain target electromagnetic (EM) responses, which demands significant efforts to investigate the enormous number of possible meta-atom structures. In this paper, a deep neural network approach is introduced that significantly improves on both speed and accuracy compared to techniques currently used to assemble metasurface-based devices. Our neural network approach overcomes three key challenges that have limited previous neural-network-based design schemes: input/output vector dimensional mismatch, accurate EM-wave phase prediction, as well as adaptation to 3-D dielectric structures, and can be generically applied to a wide variety of metasurface device designs across the entire electromagnetic spectrum. Using this new methodology, examples of neural networks capable of producing on-demand designs for metaatoms, metasurface filters, and phase-change reconfigurable metasurfaces are demonstrated.Here we propose an implicit way to construct and train the networks to predict the amplitude and phase responses of meta-structures. For a typical meta-structure, like the one shown in Fig. 1A,

Reversible Switching of Optical Phase Change Materials Using Graphene Microheaters

“…[39] and assume that any intermediate value between the crystalline and the amorphous states can be reached. [45] To optimize the meta-atom geometry, we utilized a deep neural network design framework demonstrated in Ref [79] . The neural network was set to ensure close-to-2π phase coverage and to maintain high transmittance for GSST meta-atoms constrained to a maximum thickness of 400 nm.…”

Reconfigurable all-dielectric metalens with diffraction-limited performance

“…In general, an active metasurface with j optical states (j ≥ 2) each characterized by m phase levels demands a minimum of m j distinct meta-atoms. The design problem, whose complexity escalates rapidly with increasing m and j, is best handled with deep learning based meta-atom design algorithms 63,64 and will be the subject of a follow-up paper.…”

Design for quality: reconfigurable flat optics based on active metasurfaces

“…Therefore, this method enables efficient nanophotonic device evaluation and its design prediction. Overall, ML-assisted design approaches have already shown promising progress in developing a variety of high-performance metasurfaces [310,[314][315][316][317][318].…”

“…When applied to active elements, the ML-assisted approaches were demonstrated to generate efficient designs of multifunctional and reconfigurable metadevices [316,323,325]. Encountered by the dramatically enlarged DOFs and associated uncertainties, design constraining rules, which take into account experimental imperfections, have been incorporated in several design frameworks to enhance their robustness [316,317,320]. In Section 7, we illustrate the use of ML methods to blueprint multifunctional active metadevice designs based on O-PCMs.…”

“…While the MOO techniques enable the exploration of significantly expanded design spaces, it is also highly desirable to develop design generation and validation tools in place of full-wave electromagnetic modeling to complement the optimization algorithms and enhance the design throughput. A recently emerged DNN-based metasurface design and characterization approach [316,323,328] offers a promising solution to address the design throughput challenge. In the study by An et al [316], an all-dielectric reconfigurable meta-atom design based on NNs was presented.…”