Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Data driven approaches in nanophotonics: A review of AI-enabled metadevices

Huanshu Zhang, Lei Kang, Sawyer D. Campbell, Jacob T. Young, Douglas H. Werner

TL;DR

The paper surveys data-driven strategies for AI-enabled nanophotonic metadevices, presenting a model-centric view that moves beyond brute-force EM simulations toward fast surrogate models. It covers image-based and parameter-based design, transformer- and attention-based architectures, and methods for predicting mutual coupling and ensuring fabrication-friendly robustness. Key contributions include a comprehensive catalog of DL approaches (CNNs, GANs, VAEs, diffusion models, tandem networks, and encoder-only transformers), practical cost considerations, and guidance on data efficiency and manufacturability. The review identifies promising directions such as physics-informed learning and LLM-assisted design, while acknowledging challenges in data requirements, generalization, and scalable, interpretable models for complex metastructures.

Abstract

Data-driven approaches have revolutionized the design and optimization of photonic metadevices by harnessing advanced artificial intelligence methodologies. This review takes a model-centric perspective that synthesizes emerging design strategies and delineates how traditional trial-and-error and computationally intensive electromagnetic simulations are being supplanted by deep learning frameworks that efficiently navigate expansive design spaces. We discuss artificial intelligence implementation in several metamaterial design aspects from high-degree-of-freedom design to large language model-assisted design. By addressing challenges such as transformer model implementation, fabrication limitations, and intricate mutual coupling effects, these AI-enabled strategies not only streamline the forward modeling process but also offer robust pathways for the realization of multifunctional and fabrication-friendly nanophotonic devices. This review further highlights emerging opportunities and persistent challenges, setting the stage for next-generation strategies in nanophotonic engineering.

Data driven approaches in nanophotonics: A review of AI-enabled metadevices

TL;DR

The paper surveys data-driven strategies for AI-enabled nanophotonic metadevices, presenting a model-centric view that moves beyond brute-force EM simulations toward fast surrogate models. It covers image-based and parameter-based design, transformer- and attention-based architectures, and methods for predicting mutual coupling and ensuring fabrication-friendly robustness. Key contributions include a comprehensive catalog of DL approaches (CNNs, GANs, VAEs, diffusion models, tandem networks, and encoder-only transformers), practical cost considerations, and guidance on data efficiency and manufacturability. The review identifies promising directions such as physics-informed learning and LLM-assisted design, while acknowledging challenges in data requirements, generalization, and scalable, interpretable models for complex metastructures.

Abstract

Data-driven approaches have revolutionized the design and optimization of photonic metadevices by harnessing advanced artificial intelligence methodologies. This review takes a model-centric perspective that synthesizes emerging design strategies and delineates how traditional trial-and-error and computationally intensive electromagnetic simulations are being supplanted by deep learning frameworks that efficiently navigate expansive design spaces. We discuss artificial intelligence implementation in several metamaterial design aspects from high-degree-of-freedom design to large language model-assisted design. By addressing challenges such as transformer model implementation, fabrication limitations, and intricate mutual coupling effects, these AI-enabled strategies not only streamline the forward modeling process but also offer robust pathways for the realization of multifunctional and fabrication-friendly nanophotonic devices. This review further highlights emerging opportunities and persistent challenges, setting the stage for next-generation strategies in nanophotonic engineering.

Paper Structure

This paper contains 11 sections, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. AI-assisted high-DoF metamaterial design
  3. Image-based methods
  4. Parameter-based methods
  5. Transformers and attentions applied in metamaterial design
  6. Prediction of mutual coupling effects
  7. Robust and fabrication-friendly metamaterial design
  8. Recent Advancements and Perspectives
  9. Conclusions
  10. Conflicts of interest
  11. Acknowledgements

Figures (6)

  • Figure 1: Image-based method for AI-assisted design of metasurfaces. (a) CNN network for design of high-DoF quasi-freeform dielectric metasurfaces. Reprinted with permission from pillai_leveraging_2021. Reprinted with permission from an_deep_2020. Copyright 2020, Optical Society of America. (b) GAN model for metasurface inverse design47. Reprinted with permission from campbell_explosion_2021. Copyright 2021, Wiley-VCH. (c) VAE and GA for metasurface inverse design. Reprinted with permission from park_free-form_2022. Copyright 2022, Optical Society of America. (d) The first diffusion probabilistic model for inverse design of meta-atoms. Reprinted with permission from zhang_diffusion_2023. The article is licensed under a Creative Commons Attribution 4.0 International License. https://creativecommons.org/licenses/by/4.0/ (e) The first probabilistic generative model in a tandem architecture (TGN) for the design of meta-atoms. Reprinted with permission from yang_enhancing_2025. Copyright 2025 American Chemical Society. (f) 100$\times$100 binary images were used for freeform metasurfaces forward and inverse design. Reprinted with permission from gahlmann_deep_2022. Copyright 2022 American Physical Society. https://doi.org/10.1103/PhysRevB.106.085408
  • Figure 2: Parameter-based method for AI-assisted design of metasurface. (a) An H‑shaped planar structure design using DNN. Reproduced with permission from malkiel_plasmonic_2018. TThe article is licensed under a Creative Commons Attribution 4.0 International License. https://creativecommons.org/licenses/by/4.0/. (b) Multilayered nanoparticle structures. Reproduced from peurifoy_nanophotonic_2018. © 2018 The Authors, with permission under the Creative Commons Attribution–NonCommercial 4.0 International License (CC BY-NC 4.0). http://creativecommons.org/licenses/by-nc/4.0/ (c) An all-dielectric metasurface with multiple resonant modes and near-field coupling between elements. Reprinted with permission from nadell_deep_2019. Copyright 2019, Optical Society of America. (d) A rectangular-shaped phase-modulating meta-structure. Reprinted with permission from xu_efficient_2021. Copyright 2021, Optical Society of America. (e) A 3D Born-Kuhn type chiral metasurface. Reprinted with permission from liao_deep_2022. Copyright 2022, Optical Society of America. (f) AI-assisted true 3D plasmonic high-DoF metamaterials design. Reproduced with permission from zhang_fixed-attention_2025. Copyright 2025 Optical Society of America. (g) A stacked, twisted gold split ring resonator with dielectric spacers. Reproduced with permission from ma_deep-learning-enabled_2018. Tandem networks were used to design thin-film metasurfaces. Reproduced with permission from liu_training_2018. Copyright 2018 American Chemical Society. (i) Metal-dielectric-metal periodic gap-plasmon based half-wave plate metasurface design based on biAE. Reproduced with permission from mall_fast_2020. Copyright 2020 Institute of Physics. (j) Metamaterial absorber design based on tandem networks. Reproduced with permission from hou_customized_2020. The article is licensed under a Creative Commons License. https://creativecommons.org/licenses/by/4.0/ (k) A chiral plasmonic Born–Kuhn metamaterial design based on multi-task learning. Reproduced with permission from han_neural-network-enabled_2023. Copyright 2023 American Chemical Society. (l) A dagger-shaped Ag array and an Ag mirror separated by a dielectric spacer. Reprinted with permission from luo_flexible_2024. Copyright 2024, Optical Society of America.
  • Figure 3: Transformer and self-attention for AI-assisted design of metasurfaces. (a) Transformer architecture. Reproduced from vaswani_attention_2017 with permission from Google, which grants reproduction of tables and figures for scholarly works provided proper attribution is given. (b) Encoder-only transformers for the design of broadband solar metamaterial absorbers. Reproduced from chen_broadband_2023 Copyright 2023 The Authors, Advanced Photonics Research published by Wiley-VCH GmbH, under the terms of the Creative Commons Attribution License. https://creativecommons.org/licenses/by/4.0/ (c) Dielectric metasurface design based on encoder-only transformer models. Reproduced from chen_alldielectric_2024 Copyright 2023 The Authors, Advanced Optical Materials published by Wiley-VCH GmbH (d) Improved transformer combined with a CGAN for the inverse-design of graphene terahertz multi-resonant metasurfaces. Reproduced with permission from huang_artificial_2024. Copyright 2023 IEEE. (e) GPT for inverse design of multilayer thin film structures. Reprinted with permission from ma_optogpt_2024. The article is licensed under a CC-BY 4.0 License. https://creativecommons.org/licenses/by/4.0/ (f) Mid-infrared metasurface-embedded Fabry-Perot filters design via FC layers that couple with a CNN-self-attention module. Reproduced with permission from yuan_multitask_2024. Copyright 2024 American Chemical Society.
  • Figure 4: AI-assisted design of metasurfaces with mutual coupling effects. (a) Prediction of electromagnetic responses of individual meta-atoms when mutual coupling between nonidentical neighbours is present via CNN. Reproduced from an_deep_2022 Copyright 2021 The Authors, Advanced Optical Materials published by Wiley-VCH GmbH (b) U-Net-based CNN for the modelling of complex, aperiodic plasmonic metasurfaces that can extend to arbitrarily large sizes. Reproduced with permission from majorel_deep_2022. Copyright 2022 American Chemical Society. (c) Rapid calculation and optimization of metasurfaces incorporating meta-atom interactions. Reproduced from ma_incorporating_2023. Copyright 2023 The Authors, Advanced Photonics Research published by Wiley-VCH GmbH, under the terms of the Creative Commons Attribution License. https://creativecommons.org/licenses/by/4.0/ (d) A DL optimizer for large-aperture meta-lens design via AE. Reprinted with permission from ha_physics-data-driven_2023. The article is licensed under a CC-BY 4.0 License. https://creativecommons.org/licenses/by/4.0/
  • Figure 5: AI-assisted design of robust and fabrication-friendly metasurfaces. (a) PGGAN with self-attention rapidly output freeform metasurface designs that surpass topology-optimized devices in efficiency and robustness. Reproduced with permission from wen_robust_2020. Copyright 2020 American Chemical Society. (b) A U-net based DNN with evolutionary optimization to design metasurfaces whose efficiency persists across fine-grained fabrication-induced edge deviations. Reprinted with permission from jenkins_establishing_2021. The article is licensed under a Creative Commons Attribution 4.0 International License. https://creativecommons.org/licenses/by/4.0/ (c) An end-to-end generative-modelling pipeline that learns manufacturable free-form dielectric metasurface shapes. Reproduced with permission from tanriover_deep_2022. Copyright 2022 American Chemical Society. (d) A DL-generated, fabrication-constrained library of free-form meta-atoms for the design of metasurface collimators. Reprinted with permission from ueno_dual-band_2023. The article is licensed under a Creative Commons Attribution 4.0 International License. https://creativecommons.org/licenses/by/4.0/.
  • ...and 1 more figures