All-optical diffractive neural networked terahertz hologram

Abstract

Holography has garnered an explosion of interest in tremendous applications, owing to its capability of storing amplitude and phase of light and reconstructing the full-wave information of targets. Spatial light modulators, metalenses, metasurfaces, and other devices have been explored to achieve holographic images. However, the required phase distributions for conventional holograms are generally calculated using the Gerchberg–Saxton algorithm, and the iteration is time-consuming without Fourier transform or other acceleration techniques. Few studies on designing holograms using artificial intelligence methods have been conducted. In this Letter, a three-dimensional (3D)-printed hologram for terahertz (THz) imaging based on a diffractive neural network (DNN) is proposed. Target imaging letters “THZ” with uniform field amplitude are assigned to a predefined imaging surface. Quantified phase profiles are primarily obtained by training the DNN with the target image and input field pattern. The entire training process takes only 60 s. Consequently, the hologram, that is, a two-dimensional array of dielectric posts with variational heights that store phase information, is fabricated using a 3D printer. The full-wave simulation and experimental results demonstrate the capability of the proposed hologram to achieve high-quality imaging in the THz regime. The proposed lens and design strategy may open new possibilities in display, optical-data storage, and optical encryption.

Terahertz Imaging for Breast Cancer Detection

“…Recently, numerous chemometrics methods, such as principal component analysis (PCA), partial least square method (PLS), linear discriminant analysis (LDA), decision tree (DT), genetic algorithm (GA), cluster analysis (CA), random forest (RF), partial least squares discrimination analysis (PLS-DA), ANN, and SVM, have been applied in THz imaging to improve THz data collection speed and accuracy. In addition, machine learning techniques have been used to enhance the visualization of THz images [ 81 , 82 , 83 , 84 , 85 , 86 , 87 ].…”

Research progress in optical neural networks: theory, applications and developments

“…In addition, new camera designs and optical components using D 2 NN to perform tasks can be implemented, passively used in medical technology, robotics, security, and any application that requires image and video data. For example, all-optical diffraction neural network can be used to construct holograms that can realize "THz" imaging at a very low cost through 3D printing [28], reconstructing highquality images at a high speed.…”

Transmission‐Reflection Controls and Polarization Controls of Electromagnetic Holograms by a Reconfigurable Anisotropic Digital Coding Metasurface

“…Hence, the radiation patterns of x-and y-polarized EM waves are determined by the anisotropic digital coding sequence, while the transmission and reflection modes of metasurface can be controlled in real time by the PIN diodes loaded on the middle layer. As an example, a 2-bit reconfigurable anisotropic holographic metasurface with function of focusing holographic imaging [53][54][55][56] has been proposed and designed, which is composed of 21 × 21 coding particles with a total size of 168 × 168 mm 2 (5.4λ × 5.4λ at 9.7GHz). Based on the weighted Gerchberg-Saxton algorithm, [53][54][55] the holographic images of a letter "X" for x-polarized EM waves and a letter "Y" for y-polarized EM waves are realized, respectively, and the required 2-bit anisotropic phase distributions of metasurface are shown in Figure 6a,b.…”

“…As an example, a 2‐bit reconfigurable anisotropic holographic metasurface with function of focusing holographic imaging [ 53–56 ] has been proposed and designed, which is composed of 21 × 21 coding particles with a total size of 168 × 168 mm 2 (5.4λ × 5.4λ at 9.7GHz). Based on the weighted Gerchberg–Saxton algorithm, [ 53–55 ] the holographic images of a letter “X” for x ‐polarized EM waves and a letter “Y” for y ‐polarized EM waves are realized, respectively, and the required 2‐bit anisotropic phase distributions of metasurface are shown in Figure a,b.…”