Fig. 4: Novel deep learning algorithms to drive intelligent metasurfaces.

a Tandem neural network by connecting forward and inverse neural network together, aiming to solve non-uniqueness issue in inverse design⁶³. b Synthetic neural network. The free assembly of metasurfaces in physical space corresponds to the dynamic synthesis of neural networks in algorithm space⁶⁸. c Generation-elimination network⁷¹. Such network firstly generate diverse nominations and then eliminate inferior nominations. d Reinforcement learning⁷⁴. e Transfer learning to migrate knowledge from old to new task^80,82. a Reprinted with permission from ref. ⁶³, Copyright © 2018, American Chemical Society. b Reproduced with permission from ref. ⁶⁸, Copyright © 2023, Yuetian Jia et al. c Reproduced with permission from ref. ⁷¹, Copyright © 2023, Jieting Chen et al. d Reproduced with permission from Ref. ⁷⁴, Copyright © 2019, Iman Sajedian et al. e Reproduced with permission from ref. ⁸⁰, Copyright © 2021, Ruichao Zhu et al, and ref. ⁸², Copyright © 2022, John Wiley and Sons