Fig. 1. Structure and principle of optical diffractive neural network^[31]. (a) Scheme showing the structure of deep diffractive neural networks (D²NN); (b) the light field modulation principle of the diffractive surfaces; (c) the light field modulation principle of the metasurfaces

Fig. 2. Comparison of optical paths in CGH imaging and D²NN optical information detection. (a) Holography reconstruction based on SLM^［66］; (b) optical information detection implemented by D²NN^[31]

Fig. 3. The structure of D²NNs for addressing common issues in holographic reconstruction. (a) F-D²NN achieves speckle suppression after reconstruction^[85]; (b) schematic of PC-DONN using an SLM^[87]; (c) artifact-free holographic reconstruction achieved at the output plane^[88]; (d) all-optical diffractive image denoiser^[88]

Fig. 4. Structures and functionalities of improved D²NNs. (a) Structure of Res-D²NNs^[97], where the optical residual learning module mimics the electronic Res-Net model by designing a learnable optical shortcut, which is implemented using mirrors and beam splitters between diffractive modulation layers; (b) working principle of MDA for recognizing 3D chairs, where multiple learners associated with a single view simultaneously receive the projected light field of the target^[98]; (c) P-D²NN for unidirectional image magnification and reduction^[99]

Fig. 5. Schematic diagrams of diffraction, multispectral, and multiwavelength multiplexed imaging. (a) Diffraction imager for phase information^[122]; (b) multispectral phase imager^[129]; (c) three-dimensional multiplanar phase imager via wavelength-multiplexed diffractive processing^[134]

Fig. 6. Schematic diagrams of the complex-amplitude and scattering all-optical imager. (a) Hybrid multiplexing design of all-optical complex-amplitude imaging^[135]; (b) all-optical phase imaging through a random scattering medium, and the simulated phase imaging results for unknown scatterers^[136]

Fig. 7. Schematic diagrams of solid-immersion and superoscillatory diffractive neural network imaging. (a) Solid-immersion subwavelength amplitude and phase imager and SSIM of simulated imaging results^[146]; (b) SODNN optimizing three-dimensional light field and the size of the experimental and simulated spots^[149]

Fig. 8. Progress in optical scattering processing and waveguide image transmission using diffractive neural networks. (a) Schematic diagram illustrating the principle of all-optical scattering imaging based on multi-layer diffractive neural networks, along with its reconstructed images of objects behind unknown scatterers and Pearson correlation coefficient (PCC) evaluation^[170]; (b) schematic diagram of the “vaccination” training strategy for enhancing the robustness of diffractive neural networks, and corresponding PCC comparison curves under varying interlayer displacements^[171]; (c) schematic diagram illustrating the principle of a hybrid information transmission system combining electronic encoding with diffractive decoding, and corresponding PCC comparison curves under different Fresnel numbers^[172]; (d) schematic diagram illustrating the principle of a miniaturized diffractive neural network (DN₂s) integrated at the distal facet of a multimode fiber (MMF), and the resulting reconstructed images and resolution (~4.9 µm) after overcoming modal dispersion^[58]

Fig. 9. Progress in single-pixel imaging enhanced by diffractive neural networks. (a) Schematic of task-specific image reconstruction using the diffractive network’s spectral class scores as input^[185]; (b) working principle of the all-optical hidden object/defect detection scheme^[186]; (c) the schematic drawing of a broadband single-pixel diffractive network mapping the spatial information of an input handwritten digit behind an unknown diffuser into the power spectrum at the output pixel aperture^[187]