Advanced Photonics, Volume. 7, Issue 3, 034005(2025)

Deep learning in metasurfaces: from automated design to adaptive metadevices

Yasir Saifullah1...2,3,†, Nanxuan Wu1, Huaping Wang4, Bin Zheng1,2,3, Chao Qian1,*, and Hongsheng Chen1,23,* |Show fewer author(s)
Author Affiliations
  • 1Zhejiang University, ZJU-UIUC Institute, Interdisciplinary Center for Quantum Information, State Key Laboratory of Extreme Photonics and Instrumentation, Hangzhou, China
  • 2Zhejiang University, ZJU-Hangzhou Global Science and Technology Innovation Center, Zhejiang Key Laboratory of Intelligent Electromagnetic Control and Advanced Electronic Integration, Hangzhou, China
  • 3Zhejiang University, Jinhua Institute of Zhejiang University, Jinhua, China
  • 4Zhejiang University, Institute of Marine Electronics Engineering, Ocean College, Key Laboratory of Ocean Observation-Imaging Testbed of Zhejiang Province, Hangzhou, China
  • show less
    Figures & Tables(5)
    Architecture of intelligent metasurfaces. Tunable metasurfaces are the physical foundation, and deep learning is the internal driving force. With both, intelligent metasurfaces can enable a myriad of adaptive applications without human intervention.
    Deep-learning-based metasurface design methods. (a) A traditional forward design MLP for the spectrum prediction of nanoparticles. (b) A CNN model for inverse design with customized field distributions. Two different kernel sizes can transit two different inputs to the same output shape. (c) A recurrent neural network is used to find the required series of optical responses based on the features extracted from the input nanostructure. (d) A recursive neural network that can deal with structural inputs, c represents a child node, and p represents a parent node. M represents the assembled metasurface, and m represents the metasurface component. (e) A VAE model for both inverse and forward designs between the optical responses and design geometries by setting one to both input and output reconstruction x , whereas the other one is the latent variable z before the decoder. (f) A GAN model for inverse design by generating possible patterns, approximating spectra, and evaluating accuracy. (g) Illustration of four deep transfer learning methods. (h) A deep reinforcement learning algorithm. State is the data of the detected environmental situation, and the action is the instruction to control the metasurface design. The actor network provides the design (named action in DRL), and the environment interaction provides rewards for optimization, and the next state for the next epoch training.
    Specialized algorithms in metasurface design fields. (a) Illustration of a tandem neural network, setting the EM responses as the inputs and outputs while the metasurface design serves as the middle layer. (b) Demonstration of the generation-elimination framework, from a semi-known design to an optimal solution. (c) Knowledge-inherited paradigm for a metasurface inverse design. Only the SNN network should be rebuilt, and the INN related to each child’s metasurface can be inherited. (d) Two types of physics integrated algorithms: analytical model-assisted and physical adversarial networks. The right-top loss component illustrates how numerical calculation error in PDE functions or boundary conditions (BCs) helps the efficient optimization of analytical models in the same way as normal training loss. The right-bottom physical disconfirmation, which occurs when putting the network output into the principals, helps generate adversary channels to enhance scientificity and rationality.
    Intelligent metasurfaces for invisibility cloaks and wireless communication. (a) Timeline of invisibility cloaks. (b) Schematic illustration of intelligent self-adaptive metadevices integrated with perception-decision-execution.3" target="_self" style="display: inline;">3 (c) Deep-learning-assisted self-adaptive microwave cloak. (d) Photograph of the intelligent invisible drone.154" target="_self" style="display: inline;">154 (e) Simulation results of cloaked/bare drone. Panel (c) is adapted with permission from Ref. 3, Springer Nature Limited. Panels (d) and (e) are adapted with permission from Ref. 154.
    Intelligent-metasurface-based adaptive metadevices. (a) The concept of RIS-assisted wireless communication. (b) Schematic illustration of the metasurface-based dual-channel wireless communication system. (c) A metalens for augmented reality. (d) Schematic illustration of deep-learning-based metasurface holograms. (e) Neuro-metamaterials-based object recognition system. (f) Schematic illustration of the metasurface-based LiDAR system.
    Tools

    Get Citation

    Copy Citation Text

    Yasir Saifullah, Nanxuan Wu, Huaping Wang, Bin Zheng, Chao Qian, Hongsheng Chen, "Deep learning in metasurfaces: from automated design to adaptive metadevices," Adv. Photon. 7, 034005 (2025)

    Download Citation

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

    Category: Reviews

    Received: Jan. 6, 2025

    Accepted: May. 6, 2025

    Posted: May. 6, 2025

    Published Online: Jun. 13, 2025

    The Author Email: Chao Qian (chaoq@intl.zju.edu.cn), Hongsheng Chen (hansomchen@zju.edu.cn)

    DOI:10.1117/1.AP.7.3.034005

    CSTR:32187.14.1.AP.7.3.034005

    Topics