Metasurface‐Based Terahertz Three‐Dimensional Holography Enabled by Physics‐Informed Neural Network

Artificial intelligence has revolutionized optical device design, overcoming the efficiency bottlenecks of traditional methods. For holographic metasurfaces, conventional iterative algorithms suffer from time‐consuming iterations and convergence stagnation, especially as the complexity of 3D target fields increases. While deep learning‐based algorithms have improved the trade‐off between speed and image quality, most existing models remain constrained by predefined physical scenarios. To address these challenges, we develop LM‐PINN, a physics‐informed neural network that integrates local polynomial fitting with multi‐plane wave propagation, for the rapid design of terahertz 3D holographic metasurfaces. Through self‐supervised training, LM‐PINN enables direct end‐to‐end mapping from target holographic patterns to metasurface structures within a predefined holographic configuration, without labeled datasets. Both simulated and experimental results from LM‐PINN demonstrate higher imaging quality than traditional iterative algorithms. We further incorporate distance encoding into LM‐PINN, yielding Dist‐LM‐PINN. Dist‐LM‐PINN enables a single trained model to generalize across varying diffraction distances and demonstrate 2D/3D holographic scenarios without retraining by utilizing the precomputed electric‐field distributions at the metasurface plane as inputs. The proposed framework typically completes inference in less than 1 second, providing a multifold speed advantage over traditional algorithms and offering a fast and flexible framework for high‐quality, real‐time, and large‐scale 3D holographic technologies.