Fig. 1. Micro-nano structure response predicted by neural network^[24]. (a) Schematic of 3D micro-nano structure modeling; (b) Poynting vector acquired by neural network predicting; (c) far-field scattering model acquired by neural network predicting

Fig. 2. Fully-connected neural network for inverse design of micro-nano structures^[27]. (a) Schematic of parameters of structure with multi-layer film; (b) cascading network; (c) response curves corresponding to the prediction structures of unfiltered training network and filtered training network; (e) spectral response of the prediction structure of the cascade network

Fig. 3. Research on micro-nano structure based on response prediction^[28-31]. (a) Multilayer nanoparticle structure; (b) chiral metasurface structure; (c) H-type metasurface structure; (d) electromagnetic response prediction array metasurface structure

Fig. 4. Spectrum prediction and inverse design of different structures using generating adversarial neural network^[34]. (a) Structure prediction and inverse design; (b) principle of generating adversarial neural network; (c) comparison of target structure and prediction structure; (d) comparison of spectrum response of target structure and prediction structure

Fig. 5. Multi-functional metasurface neural network system^[35]. (a)--(h) Structure diagram of neural network system; (i) metasurface system; (j)--(l) application including multifunctional beam shaping, all-optical computation, and spatial wavelength-polarization multiplexing holography, respectively

Fig. 6. Deep learning reduces the data dimension of the nano-structure^[36-37]. (a) VAE encoding the metamaterial structure and spectral decoding in latent space; (b) VAE network design on demand, reflection spectra in the 40--100 THz frequency; (c) one-to-one mapping of the structure to the response

Fig. 7. Deep learning and topology applied to the inverse design problem^[44]. (a) Schematic of micro-nano structure and spatial mapping relationship; (b) nano-structure distribution and efficiency histogram when the number of iterations is 5, 20, 50, 100, respectively

Fig. 8. Network structure and experimental results of phase restoration using deep learning^[52]

Fig. 9. Ultra-lightweight real-time 3D holography using deep learning^[56]

Fig. 10. Experimental results of high-speed hologram generation by self-coded computer generated holography^[57]

Fig. 11. DCAN model and feature maps^[63]. (a) Schematic of DCAN model; (b) feature maps of ‘boats’ test image in DCAN each-stage

Fig. 12. Deep convolutional auto-encoder network model for optimizing mask pattern^[64]

Fig. 13. Schematic of an end-to-end neural network model^[66]

Fig. 14. Structure of hybrid neural network^[82]

Fig. 15. Thick scattering media imaging using hybrid neural networks^[82]. (a) Scattering medium imaging experimental device; (b) speckle images after scattering medium; (c) reconstructed images of mixed neural network; (d) real target images

Fig. 16. Using depth neural network to achieve phase imaging by scattering^[83]. (a) Schematic of phase image reconstructed by scattering; (b) structure of the deep neural network model and its training process

Fig. 17. Training flow of microscopic imaging neural network^[84]. (a) Training of microscopic imaging neural network; (b) results of microscopic imaging neural network

Fig. 18. Diffraction deep neural network^[93]. (a) Incident wave propagation of diffraction neural network; (b) system structure of classified diffraction neural network

Fig. 19. Optical pulse shaping system with D2NN^[94]. (a) Structure of pulse shaping diffraction network; (b) numerical simulation and experimental results of pulse shaping

Fig. 20. Optical logic operation with D2NN^[95]. (a) Optical logic neural network system structure; (b) regional division of incident plane

Fig. 21. Optical decoder using single layer diffraction element^[96]. (a) Decryption of a single image; (b) decryption of a class of images

Fig. 22. Reconfigurable diffraction processor^[97]. (a) DMD, phase-type SLM, and CMOS sensors constitute the internal structure of DPU; (b) DPU is constructed as DNN structure

Fig. 23. Structure of single-pixel diffraction neural network for spectral coding classification^[98]

Fig. 24. Classification of vowels using photon neural network and physical wave neural network^[99-100]. (a) Photon neural network units on-chip; (b) acoustic neural network system