Chinese Optics Letters, Volume. 23, Issue 8, 083601(2025)

Gradient-descent optimization of metasurfaces based on one deep-enhanced RseNet

Yi Xu1, Fu Li1, Jianqiang Gu1、*, Quan Xu1, Zhen Tian1, Jiaguang Han1,3, and Weili Zhang2、**
Author Affiliations
  • 1Center for Terahertz Waves and College of Precision Instrument and Optoelectronics Engineering, State Key Laboratory of Precision Measurement Technology and Instruments, Tianjin University, Tianjin 300072, China
  • 2School of Electrical and Computer Engineering, Oklahoma State University, Stillwater OK 74078, USA
  • 3Guangxi Key Laboratory of Optoelectronic Information Processing, School of Optoelectronic Engineering, Guilin University of Electronic Science and Technology, Guilin 541004, China
  • show less
    Figures & Tables(5)
    Schematic diagram of the proposed metasurface design framework based on only one forward neural network.
    Illustration of the all-silicon meta-atom library. (a) Schematic diagram of meta-atoms for nine classes of silicon pillars. (b) Transmissive amplitude-phase spectra of 5000 randomly sampled meta-atoms of structure types 1 × 1 (top row), 1 × 2 (middle row), and 1 × 3 (bottom row) under x- and y-polarized incidences at frequencies of 0.7 and 1.0 THz, respectively.
    Framework and characterization of the deep learning model. (a) Components and hyperparameters of each layer of the neural network. (b) Training and validation loss curves for three model configurations: base, medium, and large. (c) True and predicted amplitude and phase values for two randomly selected parameter sets from the test dataset.
    Design and characterization results of the designed holographic metasurface. (a) The SEM image of the whole sample and the magnification of the selected area. (b) The intensity and phase of the incident terahertz beams scanned without samples at 1.0 THz. (c) The target, paradigm-optimized, electromagnetically simulated, and experimentally measured results of the hologram for x- (top row) and y-polarizations (bottom row), respectively.
    Design and characterization results of the designed vortex metasurface. (a) Target and paradigm-optimized phases of the metasurface for x- and y-polarizations, respectively. (b) The SEM image of the whole sample and magnification of the selected area. (c) The intensity and phase of terahertz beams scanned without samples at 0.7 THz. (d) The phase and intensity of the xoy plane at z = 8 mm, as well as the intensity of xoz and yoz sections calculated by electromagnetic simulation software for x-polarization (top row) and y-polarization (bottom row), respectively. (e) The phase and intensity of the xoy plane at z = 8.5 mm, as well as the intensity of xoz and yoz sections measured experimentally for x-polarization (top row) and y-polarization (bottom row), respectively.
    Tools

    Get Citation

    Copy Citation Text

    Yi Xu, Fu Li, Jianqiang Gu, Quan Xu, Zhen Tian, Jiaguang Han, Weili Zhang, "Gradient-descent optimization of metasurfaces based on one deep-enhanced RseNet," Chin. Opt. Lett. 23, 083601 (2025)

    Download Citation

    EndNote(RIS)BibTexPlain Text
    Save article for my favorites
    Paper Information

    Category: Nanophotonics, Metamaterials, and Plasmonics

    Received: Jan. 24, 2025

    Accepted: Apr. 3, 2025

    Published Online: Jul. 15, 2025

    The Author Email: Jianqiang Gu (gjq@tju.edu.cn), Weili Zhang (weili.zhang@okstate.edu)

    DOI:10.3788/COL202523.083601

    CSTR:32184.14.COL202523.083601

    Topics