Vision represents a fundamental channel for human information acquisition, with display technology serving as a critical medium for visual information transmission. The evolution of display technology has progressed from black-and-white to color display, enhancing both information density and visual experience. In the pursuit of more immersive visual experiences, significant advancements have been made in three-dimensional (3D) display technologies. Holographic displays, which can reconstruct the complete complex amplitude of light waves, provide comprehensive depth cues required by human vision. Consequently, these displays are regarded as one of the optimal implementations of true 3D displays, offering substantial potential applications in virtual reality (VR), augmented reality (AR), and holographic communication.Optical holography enables 3D display through recording interference patterns between object and reference waves on a recording medium. The original object wavefront can be reconstructed through diffraction when the recorded hologram is illuminated by the identical reference wave. Computer-generated hologram (CGH) simulates the diffraction process from object wave to hologram plane computationally, avoiding the complexities of optical recording. The complex amplitude on the hologram plane is encoded into a hologram using a specified reference wave and encoding algorithm. Holographic display is achieved by loading the CGH onto a spatial light modulator (SLM) to modulate incident light, thereby reconstructing the desired object wavefront. CGH offers advantages over optical holography by eliminating strict recording requirements and enabling dynamic 3D scene generation. This has established CGH as a fundamental approach for dynamic 3D holographic displays and a primary research focus in 3D display technology.However, traditional CGH algorithms encounter several challenges. The simulation of diffraction processes from object to hologram plane requires intensive computational resources. Additionally, quality optimization methods typically necessitate multiple iterations, impeding real-time generation. While CGH primarily applies to virtual 3D scenes, generating holograms of real-world objects remains problematic. Disparities between CGH and actual optical devices can compromise experimental quality. The challenge of achieving real-time, high-quality CGH generation has substantially restricted the practical implementation of holographic display technologies. These constraints have limited CGH's practical applications and remain difficult to resolve using conventional algorithms.In recent years, incorporating deep learning into CGH has produced notable advancements in both generation speed and quality. Progress has also been achieved in hardware integration and adapting CGH to various application scenarios. Deep learning enables the generation of real-time, high-quality CGH, addressing the limitations of traditional algorithms. Therefore, synthesizing existing research is essential to guide the future development of learning-based CGH.

The CGH generation algorithms based on deep learning are summarized (Fig. 3). In terms of neural network architecture optimization, researchers have enhanced the feature extraction capabilities and incorporated optical physical principles to expand the receptive field and improve the physical interpretability of neural networks (Fig. 6). Specifically, using complex-valued neural networks has improved both the speed and quality of CGH generation. Regarding dataset and loss function design, the combination of multiple loss functions and the use of depth-of-field rendering for supervision have led to enhanced quality (Fig. 7). Manually generated datasets have also reduced the difficulty of data acquisition and improved the generalization ability of neural networks. For joint hardware optimization, real optical reconstructions captured by cameras have been used to guide the network in learning the discrepancies between ideal diffraction models and real-world light propagation (Fig. 8). By leveraging deep learning, the adverse effects of practical imperfections—such as non-uniform light sources—have been significantly mitigated, leading to substantial improvements in experimental reconstruction. For real-world CGH generation, depth information is obtained using deep learning-based depth estimation techniques. And CGH can be directly generated by combining deep learning with various data acquisition methods (Fig. 9). Learning-based CGH generation enables real-time synthesis of holograms for real-world scenes, greatly promoting practical applications of holographic displays. In terms of hologram compression and adaptation, deep learning approaches have enabled both compression and decompression of holograms, and many algorithms are proposed to flexibly controlling parameters such as reconstruction size and depth (Fig. 10). In the end, existing challenges and future research directions are discussed, including improvements in performance and efficiency, enhancement of physical interpretability, and advancement toward practical applications.

Learning-based methodology has emerged as a promising approach for addressing the limitations of traditional CGH algorithms and has become a significant research focus. Current approaches demonstrate substantial improvements in generation speed and quality, while various optimization strategies continue to be explored. Additional research remains necessary to enhance learning-based CGH algorithms regarding generation efficiency, physical interpretability, and adaptability to real-world scenarios. Learning-based CGH generation algorithms represent a crucial direction for advancing CGH technology, though further development is required for practical applications.