Systematic Review of Artificial Intelligent-Driven Inverse Design for Terahertz Metamaterials

Abstract

Terahertz (THz) metamaterials, with their exceptional ability to precisely manipulate the phase, amplitude, polarization and orbital angular momentum (OAM) of electromagnetic waves, have demonstrated significant application potential across a wide range of fields. However, traditional design methodologies often rely on extensive parameter sweeps, making it challenging to address the increasingly complex and diverse application requirements. Recently, the integration of artificial intelligence (AI) techniques, particularly deep learning and optimization algorithms, has introduced new approaches for the design of THz metamaterials. This paper reviews the fundamental principles of THz metamaterials and their intelligent design methodologies, with a particular focus on the advancements in AIdriven inverse design of THz metamaterials. The AIdriven inverse design process allows for the creation of THz metamaterials with desired properties by working backward from the unit structures and array configurations of THz metamaterials, thereby accelerating the design process and reducing both computational resources and time. It examines the critical role of AI in improving both the functionality and design efficiency of THz metamaterials. Finally, we outline future research directions and technological challenges, with the goal of providing valuable insights and guidance for ongoing and future investigations.

Keywords

terahertz / metamaterils / artificial intelligence / inverse design

1 Introduction

Terahertz(THz)waves are typically defined as electromagnetic waves with frequencies ranging from \({0.1}\mathrm{{THz}}\) to \({10}\mathrm{{THz}}\) , positioned between microwave and infrared radiation in the electromagnetic spectrum. Due to the lack of efficient \(\mathrm{{THz}}\) emitters and detectors, this region has often been referred to as the “THz gap”[ 1 - 4 ]. In wireless communications, the higher carrier frequency of \(\mathrm{{THz}}\) waves compared to millimeter waves enables significantly larger channel capacity, positioning THz waves as a promising candidate for signal carriers in \(6\mathrm{G}\) and beyond wireless communication systems [ 5 - 7 ]. Moreover, the \(\mathrm{{THz}}\) spectrum encapsulates numerous molecular spectral signatures, making it highly applicable in fields such as chemical identification \(\left\lbrack {8,9}\right\rbrack\) , material characterization [ 10 ], atmospheric and astrophysical studies [ 11 , 12 ] and gas sensing [ 13 , 14 ] . Additionally, the non-ionizing nature of \(\mathrm{{THz}}\) radiation makes it an ideal tool for applications including nondestructive testing [ 15 , 16 ] , security screening [ 17 ], biomedical imaging [ 18 - 20 ] and cultural heritage preservation [ 21 ]. Natural materials generally lack a strong or efficient response in the THz frequency range, which significantly limits the development of \(\mathrm{{THz}}\) technology.

Metamaterials and metasurfaces are materials and structures that are artificially designed and fabricated. Their physical properties do not rely on the inherent characteristics of natural materials, but are instead achieved by precisely controlling their microstructures. Due to these unique properties, metamaterials and metasur-faces exhibit enormous potential for applications across various fields. In frequency ranges such as microwaves, terahertz, infrared, and visible light, they offer enhanced functionality compared to traditional materials, such as efficient beam manipulation, improved sensor performance, information transmission, and energy harvestin [ 22 - 26 ]. Metasurfaces are the two-dimensional (2D) realization of metamaterials, where the size of their unit structures is typically smaller than the operating wavelength, and electromagnetic wave properties are primarily controlled through the surface structures. Metasurfaces not only unveil the diverse physical properties of electromagnetic waves but also enable the realization of numerous technologically significant \(\mathrm{{THz}}\) devices, such as metalens [ 27 - 29 ], waveplates [ 30 - 32 ], beam steering [ 33 - 35 ]. The traditional design process for metasurfaces primarily relies on finite-difference time-domain (FDTD) simulations or finite element modeling (FEM)[ 36 - 38 ]. This approach typically involves simulating unit cells using periodic boundary conditions and subsequently assembling multiple unit cells to create large-area systems. However, such processes are often time-consuming, and the designed metasur-face units frequently fail to achieve the desired optical effects [ 39 - 42 ]. Moreover, due to the discrepancy between the boundary conditions used in array simulations and unit cell design, the inevitable coupling interactions among different metasurface units introduce unavoidable errors [ 43 ].

To address the challenges and limitations in THz metasurface design, employing artificial intelligence (AI) has emerged as a promising solution [ 44 , 45 ] . AI typically refers to programmable computers simulating human intelligence, with its implementation relying on a variety of algorithms. In this context, we explore how different algorithms contribute to the design of THz metasurfaces. To date, the algorithms applied to \(\mathrm{{THz}}\) metasurface design can be broadly categorized into two types: traditional optimization algorithms and state-of-the-art machine learning-based algorithms [ 46 - 49 ]. Inverse design represents a transformative strategy for addressing bottlenecks in THz metasur-face design [ 50 - 52 ]. Unlike traditional forward design approaches, inverse design does not rely on specific methodologies but leverages mathematical tools to solve physical problems. By defining device functionality as an objective function and applying constrained optimization, it enables optimal designs, making it particularly suited for addressing challenges where accurate models are difficult to establish. Advancements in deep learning have revolutionized the design process of metamaterials, shifting traditional forward design approaches to inverse design. This breakthrough allows for property prediction and geometry generation without requiring prior knowledge [ 50 ]. As a result, the devices obtained through inverse design are often complex, random, and non-intuitive free-form structures. By skillfully tailoring the variables, weights, and penalty terms in the loss function, inverse design can effectively handle multi-objective problems [ 53 ].

This paper investigates advanced intelligent design methodologies for \(\mathrm{{THz}}\) metasurfaces, as illustrated in Fig. 1 . The AI-driven methods offer innovative solutions for design of \(\mathrm{{THz}}\) metasur-faces, enabling precise control of electromagnetic wave parameters such as phase, amplitude, and polarization. The AI-driven inverse design of THz metamaterials has demonstrated significant potential in enhancing applications in communications, sensing, and holography. The paper is organized as follows. Section 2 introduces the intelligent design of \(\mathrm{{THz}}\) metasurfaces. Section 3 details the proposed method for intelligent electromagnetic control using \(\mathrm{{THz}}\) metasurfaces. In Section 4, we explore the applications of AI-driven inverse design techniques for THz meta-surfaces, with a particular focus on their roles in communication, holography, and sensing. Finally, Section 5 presents the conclusions and perspectives for future research.

2 Intelligent Design of THz Metama- terials

2.1 Evolutionary Optimization

Evolutionary optimization (EO) is a population-based numerical optimization algorithm that mimics the behavior of biological systems. As a global optimization algorithm, it seeks to identify the global optimum, distinguishing it from topology optimization, which often converges to local optima, and does not rely on gradient information [ 50 ]. Here, we introduce several evolutionary algorithm (EA) techniques widely used in THz metasurface design.

The genetic algorithm (GA) is a gradient-free optimization method that mimics the principles of natural selection and genetics. Its fundamental concept involves encoding potential solutions into a chromosome-like data structure [ 54 ], and employing iterative evolution to identify the optimal solution. The evolutionary process involves three key stages:(1) Selection: Choosing superior designs as parents for the next generation.(2) Crossover: Recombining pairs of parents to generate offspring, thereby increasing population diversity.(3) Mutation: Altering parts of the chromosomes to maintain genetic diversity. Each subsequent iteration of the algorithm represents a new generation of the population. An objective function is defined based on the requirements of the optimization problem, and it evaluates the fitness of individuals in the population. With each new generation, the fitness results improve compared to the previous one, driving the evolution of the population through iterations. The algorithm terminates when the fitness criteria are met or a predefined number of iterations is reached [ 55 ].

In practical applications, the GA has been refined to meet various specific requirements, resulting in numerous variants. To address the growing demand for broadband absorption, integrating THz metasurfaces onto absorber substrates has significantly enhanced absorption bandwidth. Wang et al.[ 56 ] proposed an ultra-broadband tunable THz metasurface absorber optimized through a GA and Fabry-Pérot resonance. Automation was achieved by controlling the random coding structure of the \({\mathrm{{VO}}}_{2}\) layer via the GA, followed by designing a Fabry-Perot resonant structure to expand the absorber’s bandwidth, as illustrated in Fig. 2 (a). The GA-enhanced \(\mathrm{{THz}}\) metasurface absorber achieves an absorption bandwidth range from \({1.84}\mathrm{{THz}}\) to \({8.36}\mathrm{{THz}}\) , with a relative bandwidth of up to \({127.8}\%\) , and the optimized absorption rate exceeds \({90}\%\) . By adjusting the conductivity of \({\mathrm{{VO}}}_{2}\) through ambient temperature, the maximum absorption rate could be tuned from 4% to 100%.

However, with the advancement of \(\mathrm{{THz}}\) metasurfaces and the increasing demand for enhanced functionalities, the design objectives for THz metasurfaces have become increasingly stringent, evolving from single-objective optimization to multi-objective optimization problems. In most multi-objective optimization scenarios, when certain design objectives are prioritized over others, a common approach is to assign higher weights to the higher-priority objectives within the objective function. However, using fixed weights may cause the optimization process to fail to adapt to the complex relationships and priority changes between objectives, thereby affecting the effectiveness and efficiency of the optimization. Therefore, using fixed weights for a single optimization often does not yield satisfactory results, and a more flexible weight adjustment strategy is needed to improve the quality of optimization and convergence speed [ 58 ]. Even assigning very small weights to less critical objectives can cause the GA to deviate and significantly slow its convergence. To address this, Samad et al.[ 57 ] proposed an adaptive genetic algorithm (AGA) to tackle multi-objective optimization problems. This technique allows the GA to first converge on a generation of individuals that satisfactorily meet high-priority sub-objectives before attempting to improve lower-priority sub-objectives. This iterative process continues until the GA termination criteria are satisfied, as illustrated in Fig. 2 (b). The advantages of the AGA were demonstrated by addressing four specific challenges. Fig. 2 (b) shows the following optimizations: a binary-mode plasmonic reflective array for beam control, anasymmetric leaky-wave antenna array radiating TE and TM waveguide modes in any direction, a compact birefringent all-dielectric THz metasurface for finer lateral resolution, and a visible-light-transparent infrared emitter/absorber device combining materials like ITO and fused silica for solar cell cooling.

Particle swarm optimization (PSO) is a gradient-free method developed by simulating the social behavior of bird flocks. It is particularly well-suited for design problems with a large number of parameters and unknown optimal starting points. The optimization process begins with a swarm of particles randomly initialized in the search space. These particles search for the global optimum through both individual exploration and collective cooperation. The position update of each particle during the optimization iterations depends on its momentum, the best personal position, and the best global position of the swarm. Ultimately, the particles converge to a common optimal position, representing the global optimum [ 59 ]. The optimization objective serves as the evaluation criterion for PSO, with common goals including maximizing efficiency and phase-amplitude adaptability.

Thompson et al.[ 60 ] proposed a polymer-embedded broadband THz metasurface reflector based on PSO. In their study, PSO was used to discover various broadband reflector designs, aiming to find structures with a reflectance greater than 99% over a wavelength range fluctuating by \(\pm {20}\%\) around \({1.55\mu }\mathrm{m}\) . Ultimately, three potential polymer-embedded THz metasur-face designs were identified, each offering distinct performance advantages, such as ultra-wideband reflectance and polarization independence, as shown in Fig. 3 (a). However, the conventional PSO optimization algorithm is often prone to getting trapped in local optima, making it challenging to navigate the vast parameter space of multi-objective THz metasurface optimization. The interdependencies between objectives further complicate the search for an optimal solution. Therefore, finding methods that can accelerate unit topology design while simultaneously optimizing and balancing multiple objective parameters remains a significant challenge. Xing et al.[ 61 ] proposed a method that combines the multi-objective particle swarm optimization (MOPSO) algorithm with Python-CST co-simulation, as shown in Fig. 3 (b). The design goal in the study is the reduction of radar cross-section (RCS) in a dual-band frequency range. The MOPSO algorithm helps avoid the tedious manual adjustment of structural parameters while ensuring that the maximum absorption conversion rate (ACR) bandwidth in both frequency bands exceeds 0.9 . In addition, by utilizing CST-Python collaborative simulation, this method can quickly identify unit cells that simultaneously incorporate absorption and polarization conversion mechanisms. It effectively reduces the mono-static/bi-static RCS by over \({10}\mathrm{\;{dB}}\) within the dual-band frequency range of \({2.07}- {3.02}\mathrm{{THz}}\) and 3.78-4.71 THz. Building upon the MOPSO algorithm, Ma et al.[ 62 ] proposed a constrained multi-objective optimization (CMOP) model based on the multi-objective PSO algorithm, as shown in Fig. 3 (c). After optimization using the proposed CMOP model, three discrete narrow resonance peaks were achieved within the \({0.5}- {2.0}\mathrm{{THz}}\) range, with absorption rates of \({99.1}\%,{90.0}\%\) , and 99.9 \(\%\) , respectively. Compared to traditional brute-force methods, the reflection loss of all resonant modes shows significant reduction.

2.2 Topology Optimization

The goal of topological optimization is to employ mathematical tools to determine the optimal structural configuration. In recent years, it has emerged as a crucial tool for the inverse design of THz metasurfaces. The process of topological optimization for \(\mathrm{{THz}}\) metasurfaces can be described as follows: Initially, the structure and its related parameters are defined as \(X\) . Mathematical tools are then used to calculate the electromagnetic response of the structure. The desired electromagnetic response is compared with the current response to formulate a loss function \(Y\) . Gradient-based methods, such as automatic differentiation, are then employed to compute the gradient of the loss function with respect to the parameters \(X\) . This gradient information is used to update the parameters \(X\) , minimizing the value of the loss function. This process is iteratively repeated until the loss function converges to a minimum value. The final output is a set of optimized parameters \(X\) that corresponds to the desired structure. Topology optimization designs the size and shape of structures within a given space and can also be used to improve precision [ 63 ]. It considers physical laws, optimization objectives, and constraints, and uses optimization algorithms to find the best possible design within the specified limits. When the physical model used in the optimization process is accurate, this method achieves the highest precision. Moreover, the number of simulations required for topology optimization does not increase with the number of elements in the system. Therefore, in cases of high degrees of freedom, where the design involves free-form \(\mathrm{{THz}}\) metasurfaces, topology optimization can demonstrate significant advantages and high efficiency.

Topology optimization plays a crucial role in various aspects of \(\mathrm{{THz}}\) metasurface design. For example, by utilizing physical topology techniques, special functionalities can be achieved by altering the directional properties of the target material. Cui et al.[ 64 ] proposed a design concept that uses physical topology techniques to transform the sensitivity of structural unit cells into independent sensitive directions of the topological unit cells. This approach enabled the simple design of flexible metamaterial absorbers with polarization-sensitive isotropy in the THz frequency range. For a single anisotropic unit cell, topology optimization compensates by arranging four unit cells in a counterclockwise combination, resulting in diagonal symmetry, as shown in Fig. 4 (a). Under standard TE and TM polarized plane electromagnetic waves, with the incident direction perpendicular to the surface of the structure, the absorber THz metasurfaces can achieve polarization insensitivity along both the \(x\) - and \(y\) -axes of the unit cells. This result indicates that topology optimization compensation has created polarization anisotropy in the structure and successfully transitioned it to isotropy. Subsequent validation of the designed device demonstrated polarization insensitivity, polarization isotropy, wide incident angle tolerance, and high absorption performance at two distinct resonant frequencies. Specifically, the absorption rate exceeded \({94}\%\) at each resonant frequency of \({0.136}\mathrm{{THz}}\) and \({0.143}\mathrm{{THz}}\) , with isotropic polarization insensitivity across the entire incident direction and a polarization angle independence exceeding \({45}^{\circ }\) .

Topology optimization can also be used to manipulate the wavefront of \(\mathrm{{THz}}\) waves through structural optimization. Nishjima et al.[ 48 ] utilized topology optimization to design the effective refractive index distribution of a broadband THz collimator without relying on dielectric materials. The study demonstrated a technique for guiding THz wavefronts over a wide bandwidth range, effectively controlling THz wave propagation without the need for dielectric materials. This was achieved by compensating for the significant frequency dispersion. This compensation was accomplished through the spatial modulation of the effective refractive index distribution, which was controlled between parallel metal plates operating in the fundamental TE mode. As shown in Fig. 4 (b), the technique employs a topology optimization method based on the method of moving asymptotes (MMA) to design a surface profile capable of converting a point source into a planar wavefront with a width of approximately \({15}\mathrm{\;{mm}}\) . This design operates over a frequency range from 250 to \({300}\mathrm{{GHz}}\) and is realized within a compact \({20}\mathrm{\;{mm}}\times {20}\mathrm{\;{mm}}\) area. Additionally, by adjusting the objective function, various wavefront shapes can be achieved. The proposed method offers a versatile strategy for integrated THz systems. Topology optimization in THz metasurfaces is not limited to the design of unit cells alone. Rouhi et al.[ 65 ]developed a metasurface by optimizing the chemical potential of a special material, graphene, to achieve real-time wave manipulation and energy suppression through a combination of absorption and diffusion, as shown in Fig. 4 (c). This approach allows for control over amplitude, phase, or polarization. The THz metasurface designed based on this method exhibited low scattering and polarization-insensitive responses with values below \(-{10}\mathrm{\;{dB}}\) over a wide frequency band from \({1.02}\mathrm{{THz}}\) to \({2.82}\mathrm{{THz}}\) . This design successfully combines THz spectral absorption with diffusion functionality.

2.3 Machine Learning

Compared to traditional omization algorithms, machine learning can predict unknown problems by learning the complex relationships between model variables and optical properties from large datasets of known information. Machine learning has also been widely applied in electromagnetic metamaterial design, demonstrating significant potential for future advancements in this field [ 66 - 68 ]. This approach significantly reduces the computational time required for \(\mathrm{{THz}}\) metasur-face design by providing a more comprehensive and systematic optimization of the THz metasur-face properties.

The working principle of machine learning (ML) is to extract useful information from raw databases and output independent decisions to perform tasks. The time-consuming and inefficient optimization process in traditional design workflows is relatively simple and efficient for ML. Furthermore, ML can effectively reduce design biases caused by human differences in traditional design processes, offering high reliability, stability, and accuracy. Deep neural networks (DNNs), a subset of machine learning, are algorithms inspired by the biological neural networks found in nature [ 69 ]. The machine learning-based metasurface design process typically follows these steps: For a simple structure, a forward-solving algorithm is applied using a combination of various parameters as inputs to obtain the electromagnetic (EM) response dataset. This dataset is then used to train a deep neural network (DNN), which can compute the EM response when provided with new input parameters. Through the same training process, an inverse network can also be obtained. The key difference with the inverse network is that the input consists of the desired response, while the output is the geometric parameters of the structure. The forward-solving algorithm can then be used to evaluate the optimized solution to determine if the response is acceptable. Typically, the optimal solution is obtained through the cross-solving of both forward and inverse networks [ 70 ]. This section focuses on deep learning models, exploring how advanced algorithms can generate \(\mathrm{{THz}}\) metasur-faces on demand.

In recent years, deep learning has become a frontier in the THz metasurface field, overcoming many challenges that traditional design methods could not address. For example, in all-dielectric THz metasurface (ADM) design, the geometric configurations of the THz metasurfaces increase rapidly as the number of unit cells grows, leading to a much largeFr number of candidate geometries to explore. The simulation time required to accurately model the complexity of the unit cell’s ADM further exacerbates this issue, meaning that only a small fraction of possible designs can be considered. Christian et al.[ 71 ] proposed a deep learning-based method to accelerate the design of ADM. They developed a deep neural network that handles both forward and inverse design processes. In the forward design, given a specific geometry \(x\) , the corresponding S-parameters s are calculated. In the inverse design, for a specified set of desired S-parameters \(s\) , the network determines the corresponding geometry \(x\) . The performance was quantified using mean squared error (MSE) to assess whether the model had learned the underlying physics or was merely a good model. Fig. 5 (a) and (b) illustrate this approach. The network is capable of accurately predicting the frequency-dependent transmission of all-dielectric THz metasurfaces given a set of eight geometric parameters. By precomputing certain factors, the approach reduces the amount of data required for the deep network to design \(\mathrm{{THz}}\) metasurfaces with good performance. This includes using derived geometric parameters as inputs, allowing the forward model to be solved using only \({0.0022}\%\) of all possible configurations. Additionally, the authors introduced the concept of fast forward design synthesis (FFDS) as a solution to the inverse material design problem, overcoming the challenging one-to-many mapping issue. This significantly enhances the feasibility of applying deep learning in THz metasurface design. The DNN established in this study is highly accurate, achieving an average MSE of \({1.16}\times {10}^{-3}\) , with \({99}\%\) of the data having an MSE less than \({6.2}\times\) \({10}^{-3}\) . Moreover, it is five orders of magnitude faster than traditional electromagnetic simulation software, effectively addressing the complexity of designing all-dielectric THz metasurface unit cells and significantly reducing simulation time.

Chiral THz metasurfaces play a crucial role in various \(\mathrm{{THz}}\) metasurface applications, as they enable effective polarization and chirality control for incident \(\mathrm{{THz}}\) waves simultaneously. However, challenges remain in the design of chiral \(\mathrm{{THz}}\) metasurfaces that traditional design methods struggle to overcome. Typically, the design of chiral \(\mathrm{{THz}}\) metasurfaces involves two main frameworks: the first involves forward prediction of electromagnetic responses based on given \(\mathrm{{THz}}\) metasurface structural parameters, and the inverse design process retrieves structural parameters based on the desired electromagnetic response. The second framework guides the final design of chiral THz metamaterials by examining the differences in the off-diagonal elements of the Jones transfer matrix [ 72 ]. However, quantitatively designing \(\mathrm{{THz}}\) metamaterial structures to achieve the desired asymmetric transmission (AT) phenomenon, or quickly predicting the \(\mathrm{{THz}}\) response as the structure varies, remains a significant challenge. Whether using traditional methods that rely on direct physical principles and empirical trial-and-error, which waste computational resources and time, or random algorithm-based optimization, the time cost increases exponentially as the design scale grows. Therefore, improving the design efficiency of chiral \(\mathrm{{THz}}\) metasurfaces has become one of the most difficult challenges to overcome.

Gao et al.[ 73 ] proposed a deep learning-based approach to accelerate the design of chiral \(\mathrm{{THz}}\) metamaterials. As shown in Fig. 5 (c), a “G”-shaped chiral THz metamaterial structure was used as a prototype to generate a large number of samples for the model. The deep learning framework, shown in Fig. 5 (d), consists of two bidirectional networks: the spectral network (SN) and the extension network (EN). In the SN, the forward path includes the tensor down-sampling (TDS) module and the Tconv upsam-pling (TUS) module, which convert structural parameters into response spectra. The reverse path effectively solves the many-to-one problem in THz metamaterial design, with the forward path of the trained SN network assisting in this process. In the EN, the forward path enables high-precision prediction of the THz metamate-rial’s AT spectrum, while the reverse path directly retrieves the structural parameters from the desired AT response. This allows the model to autonomously decode the non-intuitive relationship between the chiral \(\mathrm{{THz}}\) metamaterial structure and its corresponding electromagnetic response. Both networks can perform bidirectional tasks, enabling both forward prediction and inverse design. The results indicate that the designed model is not only applicable to the AT characteristics of chiral \(\mathrm{{THz}}\) metasurfaces but can also be extended to the design of other types of THz devices.

Although DNN-based THz metasurfaces have achieved a variety of functionalities, research on multifunctional \(\mathrm{{THz}}\) metasurfaces remains limited, with the relatively simple structures used in inverse design constraining the feasibility of multifunctionality. To address the diverse application scenarios of \(\mathrm{{THz}}\) metasur-faces, Cheng et al. employed convolutional neural networks (CNN) to perform inverse design of \(\mathrm{{THz}}\) metasurface unit cells based on target phase spectra [ 74 ]. They discussed the accuracy of CNN in the output layer, utilizing both classification and regression models. In this model, a dielectric \(\mathrm{{THz}}\) metasurface composed of pixelated unit cells is used to collect training data. Based on this dataset, a CNN model, as shown in Fig. 5 (e), is constructed. The input consists of the real and imaginary parts of the spectral data of the unit cells, as depicted in Fig. 5 (f). The operations in the hidden layers include convolution, deconvolution, and fully connected layers. The input data is reshaped into vectors and fed into the fully connected layers to learn the electromagnetic correlations within the frequency range of interest. The output is a multi-channel matrix, followed by six sets of convolution and deconvolution layers for feature extraction. Once properly trained, the neural network can identify unit cells that exhibit the desired phase response relative to frequency and polarization. After a single training session, the same model can be used to reversely design different \(\mathrm{{THz}}\) metasurfaces for holographic imaging at varying frequencies. The design is subsequently validated through full-wave simulations, demonstrating the use of a single THz metasurface for polarization multiplexing and dual-function holographic imaging. Intelligent design methodologies have significantly transformed the landscape of \(\mathrm{{THz}}\) metasurface research. Looking ahead, it is anticipated that intelligent design will unlock even greater possibilities by incorporating more geometric and physical degrees of freedom.

3 THz Metasurface for Intelligent Electromagnetic Control

THz metasurfaces represent a novel class of electromagnetic materials that enable precise control over \(\mathrm{{THz}}\) wave propagation characteristics through subwavelength-scale structural design. Compared to traditional electromagnetic wave manipulation techniques, metasurfaces offer ultrathin structures and a high degree of design flexibility, allowing for efficient modulation of various parameters such as phase, phase shifts, and orbital angular momentum. The THz spectrum finds widespread applications in communication, imaging, security scanning, and biomedical fields, with the introduction of metasurfaces further expanding the potential of these areas.

However, designing THz metasurfaces for specific functions often involves complex physical mechanisms and optimization across high-dimensional parameter spaces, which makes traditional design methods inefficient. AI-based inverse design offers an innovative approach to address this challenge [ 22 ]. Through techniques such as deep learning, generative adversarial networks (GANs), and reinforcement learning, AI-driven inverse design can autonomously generate metasurface structures optimized for predefined functional objectives. This integration not only enhances design efficiency but also uncovers innovative designs that are difficult to achieve with traditional methods.

The research explores various types of AI-driven inverse design for \(\mathrm{{THz}}\) metasurfaces, including amplitude-regulated \(\mathrm{{THz}}\) metasurface, phase modulation THz metasurfaces, polarization-controlled THz metasurface and vortex wave THz metasurface. These diverse types of THz metasurfaces employ unique design strategies to achieve precise control over \(\mathrm{{THz}}\) wave properties in specific application scenarios, thereby driving the further advancement of this technology.

3.1 AI-Driven Inverse Design of Amplitude-Con- trolled THz Metasurface

THz amplitude-controlled metasurfaces are special metasurface structures capable of controlling the amplitude of THz waves. By designing sub-wavelength structural elements, these THz meta-surfaces alter the absorption, reflection, or transmission coefficients of incident electromagnetic waves at different spatial positions, thereby enabling the modulation of the electromagnetic wave’s amplitude. THz metasurfaces designed using traditional methods still lack flexibility and diversity. THz amplitude-modulating metasur-faces rely on precise control of the electromagnetic wave’s amplitude, typically achieved by adjusting the surface structure’s geometry, material properties, or electromagnetic response. Through such design, the metasurface can modulate specific amplitudes within the THz frequency range. Amplitude-modulating THz meta-surfaces are generally composed of high-loss materials or materials with specific dielectric constants and conductivity. These materials can enhance or suppress certain parts of the electromagnetic wave, thereby altering the amplitude of the incident wave. The amplitude modulation exhibits complex nonlinear relationships between the THz metasurface’s geometric parameters and material properties. The high-dimensional and non-convex nature of the design space increases the difficulty of optimization. Moreover, each adjustment of the parameters requires performance validation, and the simulation for multiunit cell, wide-frequency designs is particularly time-consuming, which undermines optimization efficiency. In contrast, inverse design has gained widespread attention due to its flexibility and robustness in photonic applications [ 75 ]. This provides an excellent opportunity for the design of THz metasurfaces and the development of multifunctional, high-performance THz devices. In this section, we summarize the amplitude modulation of THz metasurfaces based on different inverse design methods.

Yue Wang et al.[ 76 ] proposed a machine learning-assisted bidirectional integrated learning framework, as shown in Fig. 6 (a), for designing a composite metamaterial absorber in the \({0.3}- {2.0}\mathrm{{THz}}\) range. This framework not only effectively predicts the absorption spectrum in a forward manner but also retrieves the structural parameters of the composite material from a given spectrum. Liu et al.[ 77 ] introduced the application of GANs in the inverse design of THz metamaterials. This model can directly generate the corresponding \(\mathrm{{THz}}\) metamaterial structure from the target spectral features, achieving a direct mapping from functional requirements to physical implementation, as shown in Fig. 6 (b). He et al.[ 78 ] proposed an ultrafast all-optical \(\mathrm{{THz}}\) modulation based on inverse design of \(\mathrm{{THz}}\) metasurfaces, as shown in Fig. 6 (c). By combining PSO with the FDTD method, they constructed a THz metasurface supported by electromagnetically induced transparency (EIT) effects. This design was experimentally verified for ultrafast EIT modulation at the picosecond scale. The involved THz metasurface utilized optical excitation to modulate the EIT resonance amplitude and enable slow-light switching.

3.2 AI-Driven Inverse Design of Phase-Con- trolled THz Metasurfaces

THz phase-controlled metasurfaces are functional structures that use subwavelength-scale microstructural elements to precisely control the phase of electromagnetic waves, enabling THz wave modulation. The core principle is based on phase gradients, achieved by designing microstructure unit cells with varying geometric parameters, which allows for precise control over the phase distribution of incident electromagnetic waves. This enables the modulation of THz wave propagation direction, focusing, reflection, and other characteristics. Commonly used materials, such as metals, dielectric materials, and two-dimensional materials, exhibit excellent electromagnetic response characteristics in the THz frequency range and allow for full-phase modulation in the range of 0 to \({2\pi }\) . Through periodic structural design, the metasurface can further enhance the flexible control of \(\mathrm{{THz}}\) wavefronts.

\(\mathrm{{THz}}\) phase-modulated metasurfaces have demonstrated extensive applications in fields such as phase-gradient \(\mathrm{{THz}}\) metasurfaces, \(\mathrm{{THz}}\) metasurface lenses, and \(\mathrm{{THz}}\) holographic meta-surfaces. Phase-gradient \(\mathrm{{THz}}\) metasurfaces achieve precise control over the reflection or transmission angles of \(\mathrm{{THz}}\) waves by introducing linear or nonlinear phase gradients. THz meta-surface lenses simulate the focusing and diverging functions of traditional lenses by imparting specific phase delays to incident waves at different positions. THz holographic metasurfaces, on the other hand, control phase distributions to enable holographic imaging, allowing the reconstruction of three-dimensional images or complex optical wavefronts. These studies have significantly advanced the development of THz wavefront manipulation technology, providing new technical support for imaging and information processing fields.

Zhu et al.[ 51 ] proposed a transfer learning-based method for the inverse design of functional \(\mathrm{{THz}}\) metasurfaces, treating meta-atoms as images to rapidly predict phase patterns with high accuracy. This approach significantly reduces training data requirements and design time, enabling efficient design of \(\mathrm{{THz}}\) metasur-faces with customized phase profiles for \(\mathrm{{THz}}\) applications, as shown in Fig. 7 . Hou et al.[ 79 ] proposed a deep learning-based inverse design method for all-silicon THz chiral metasurfaces, named the target-driven conditional generative network (TCGN). The proposed method can generate \(\mathrm{{THz}}\) metasurface structures that meet the required circular dichroism, and the designed all-silicon \(\mathrm{{THz}}\) metasurface can be used in phase control devices. Teng et al.[ 45 ] introduced a forward prediction and inverse design method for \(\mathrm{{THz}}\) random metasurfaces, based on deep \(\mathrm{{CNN}}\) and GA. By combining multi-objective optimization with the genetic algorithm, the design efficiency was further enhanced. The inverse design model can generate \(\mathrm{{THz}}\) metasurface structures within 10 minutes based on the target response.

3.3 AI-Driven Inverse Design of Polarization- Controlled THz Metasurfaces

THz polarization-modulated metasurfaces represent an advanced structure capable of effectively controlling the polarization state of incident \(\mathrm{{THz}}\) waves. By altering the direction of the electric field oscillation, polarization modulation enables more precise control of electromagnetic waves in various applications, including information transmission, imaging, and sensing. In the THz frequency range, polarization-modulated metasur-faces achieve flexible manipulation of the incident wave’s polarization state through the design of complex subwavelength structures. These metasurfaces primarily control the polarization state of \(\mathrm{{THz}}\) waves by designing specific sub-wavelength unit cells. Through these microstruc-tures, the \(\mathrm{{THz}}\) metasurface can alter the electric field direction of the electromagnetic wave, thereby changing the polarization state of the incident wave. Traditional methods for designing polarization-converting \(\mathrm{{THz}}\) metasurfaces require extensive numerical simulations and iterative trial-and-error processes, which are computationally intensive and time-consuming. As a result, there is a demand for faster and more efficient design approaches. The introduction of inverse design for \(\mathrm{{THz}}\) polarization metasurfaces provides a solution to these challenges. Liu et al.[ 80 ] introduced the application of GANs in the inverse design of \(\mathrm{{THz}}\) metamaterials. This model directly generates the corresponding \(\mathrm{{THz}}\) meta-material structure from the target spectral characteristics as shown in Fig. 8 (a), achieving a direct mapping from functional requirements to physical realization. Mao et al.[ 81 ] proposed a transformer-based inverse design method for THz polarization conversion devices. This method allows for the precise design of the corresponding \(\mathrm{{THz}}\) metasurfaces based on target spectra. Using this approach, two ultra-wideband polarization conversion devices were designed within the \({0.5}- {4.0}\mathrm{{THz}}\) target frequency range, with a design time of only approximately \({0.006}\mathrm{\;s}\) as shown in Fig. 8 (b). The polarization conversion ratio surpassed \({90}\%\) within the frequency range of \({0.5}- {4.0}\mathrm{{THz}}\) , and the average cross-polarization was also greater than \({50}\%\) , demonstrating excellent polarization conversion performance. Conventional inverse design methods often require a substantial amount of time to gather large datasets for training. When the input performance data accounts only for the required performance while neglecting hidden or unknown factors, one inverse design problem input may correspond to multiple solution design outputs. The one-to-one mapping between inputs and outputs in machine learning is broken, increasing the complexity of learning and impeding the discovery of stable, effective mapping rules. Moreover, during the design process involving multiple parameters, there is still room for improvement in neural networks when it comes to addressing the coupling issues between these parameters.

Lv et al.[ 82 ] proposed an improved inverse design method by applying a residual network (ResNet) to the design of dielectric material \(\mathrm{{THz}}\) metasurfaces. They designed a silicon circular split-ring structure on a silicon dioxide substrate, which can control right-handed circularly polarized (RCP) or left-handed circularly polarized (LCP) light. The ResNet architecture significantly alleviates the coupling issues between two structural parameters. Through validation, the method was shown to enable the inverse design of unit structures with arbitrary combinations of light field amplitude and phase within a specified error range as shown in Fig. 8 (c). The approach demonstrated effective parameter coupling management and exhibited broad applicability.

3.4 AI-Driven Inverse Design of Vortex Wave \(\mathbf{{THzMetasurface}}\)

Electromagnetic waves can carry two types of angular momentum: spin angular momentum (SAM) and orbital angular momentum (OAM). SAM is typically associated with circularly polarized (CP) beams and has only two values, \(\pm \hslash\) , whereas OAM can take multiple values of \(\hslash\) . Therefore, electromagnetic waves with OAM, defined as vortex beams, can carry an unprecedented amount of data, enabling high-capacity data transmission.

Traditional methods for designing \(\mathrm{{THz}}\) vortex beam metasurfaces face challenges such as low precision and complex, time-consuming design processes. In this context, we discuss the application of AI for the inverse design of \(\mathrm{{THz}}\) vortex beam metasurfaces. Chen et al.[ 83 ] proposed a method for inverting the design of chiral \(\mathrm{{THz}}\) metasurfaces based on a given target circular dichroism (CD) value. By leveraging deep learning, they achieved the inverse design of the \(\mathrm{{THz}}\) metasurface, which can modulate the phase of right-handed circularly polarized (RCP) waves at a frequency of \({1.126}\mathrm{{THz}}\) through pancharat-nam-berry (PB) phase. control. This enables the focusing of vortex beams in the THz metasurface array. This design approach not only saves time but also improves accuracy as shown in Fig. 9 (a), effectively avoiding the loss of optimal structures that may occur during traditional parameter scanning processes.

However, the aforementioned inverse design THz metasurfaces remain limited to static manipulation and are not adaptable to the constantly changing demands of various applications, lacking dynamic control. To address this, Luan et al.[ 84 ] proposed a multifunctional reconfigurable Dirac semimetal-coded metasurface based on a THz genetic algorithm. By using a genetic algorithm to inverse-design the THz metasurface coding array and adjusting the Fermi level of the Dirac semimetal, they demonstrated the ability to control the beam. This approach enables the generation of single vortex beams with \(\pm 1\) and \(\pm 2\) topological charges within the THz range, with mode purity exceeding 60%. Additionally, vortex phase convolution allows for the realization of single, double, and triple vortex beams with controllable pitch and 360-degree azimuthal angles. The system also supports beam deflection and RCS reduction, achieving a multifunctional reconfigurable THz metasurface. Peng et al.[ 85 ]extended the application of deep learning in \(\mathrm{{THz}}\) metamaterial design to the generation of multifunctional vortex beams. As shown in Fig. 9 (b), this work highlighted the efficiency of artificial intelligence-driven inverse design in addressing multi-objective and highly customized designs, demonstrating their broad applicability in achieving complex optical functionalities.

4 AI-Driven Inverse Design of Meta- surface for \(\mathrm{{THz}}\) Applications

4.1 Communication

As interest grows in \(\mathrm{{THz}}\) frequencies as a potential communication band for the future, \(\mathrm{{THz}}\) optical devices are being explored as key components for next-generation wireless communication systems. The introduction of intelligent design methods saves considerable time and effort, reducing the need for manual adjustments of various parameters. THz metasurfaces designed with intelligent techniques have already been extensively studied in the field of communication. In communication systems, precise and low-loss multi-user communication is crucial, making efficient \(\mathrm{{THz}}\) beamforming devices essential. These devices play a vital role in strengthening energy-efficient connectivity, compensating for path loss, optimizing resource usage, and enhancing spectral efficiency.

Tan et al.[ 86 ] developed an intelligent adaptive beamforming scheme achieved through deep reinforcement learning. The proposed deep learning model utilizes a fully connected neural network to predict the phase distribution required for beamforming, thereby forming the desired intensity pattern. This approach is well-suited for implementing dynamic responses to real-time changes in user locations within multi-user multiple input multiple output (MIMO) systems, as shown in Fig. 10 . The scheme employs automatic differentiation to compute the error gradients and adaptively optimizes the neural network parameters by minimizing the discrepancy between the input and predicted intensity patterns. In this approach, the predicted intensity pattern is derived from the output phase distribution. Since the model is trained by comparing the input and predicted intensity patterns, there is no need for paired input-output data of intensity and phase, unlike traditional neural networks that require large databases for training. Instead, the neural network in this scheme is trained in real-time using small-batch data from an intensifying database of intensity patterns, which grows with the historical user location data. As a result, the prediction accuracy of the neural network can adaptively improve during the training process, enabling real-time beam-forming without being constrained by pre-trained forward networks. Using this model, silicon THz metasurfaces were designed, and experimental implementation of \(2\mathrm{D}\) beamforming for \(\mathrm{{THz}}\) radiation targeting multiple spatially separated users was achieved.

4.2 Holography

Traditional optical holography metasurfaces have expanded the capacity of holography, offering the potential for revolutionary advancements in realistic imaging, information storage, and encryption [ 87 ], In recent years, significant progress has been made in this field, including achievements in multiplexing and full-color holography [ 88 , 89 ] . However, realizing more complex holographic functions or higher-quality holographic images requires precise design of metasurface arrays, which presents a significant challenge. Intelligent design methods have overcome these limitations and simplified the design process.

Hou et al.[ 90 ] proposed a deep learning approach for on-demand design of holographic THz metasurface structures, as shown in Fig. 11 (a). The proposed network architecture is inspired by the concept of conditional generative adversarial networks (cGAN), specifically tailored for \(\mathrm{{THz}}\) metasurface design, treating it as a regression problem rather than a classification task. By eliminating the adversarial process and the authenticity judgment, the complexity during training is reduced. A dedicated sub-network, referred to as the simulator, is introduced to model the optical response of the THz metasur-face. This leads to the formation of a conditional generative network (CGN), as depicted in Fig. 11 (b). In the CGN, the simulator, upon receiving structural parameters as input, can predict the amplitude and phase of the cross-polarized components. A well-trained simulator can effectively replace the time-consuming simulation process and be used to determine the predicted results of the generator. On the other hand, the generator can reverse-engineer the structural parameters based on the desired amplitude and phase of the cross-polarized components. The designed all-silicon THz metasurface not only significantly enhances the design efficiency of holographic \(\mathrm{{THz}}\) metasurfaces but also enables continuous modulation of both phase and amplitude. This greatly improves both the design efficiency and imaging quality of the holographic \(\mathrm{{THz}}\) metasurface. The resulting field distribution from the network measurements yields a mean absolute error (MAE) of 0.0148 compared to the expected field distribution.

Compared to holograms that only contain amplitude or phase information, holograms that reconstruct the lateral profile of light using complex amplitude information demonstrate superior performance, with an improved signal-to-noise ratio. Wei et al.[ 91 ] proposed an unsupervised, physics-driven deep neural network that uses unit cells with incomplete modulation of both phase and amplitude to design complex amplitude holograms based on THz metasurfaces. This method operates as a generative adversarial model, where the neural network module serves as the generator and the physical propagation module acts as the discriminator. Silicon nanocylinders with phase modulation for propagation are employed as the unit cells. By appropriately modifying the geometric parameters of the cylinders, the effective refractive index of each cylindrical waveguide can be individually manipulated to shape the wavefront and control the intensity, as shown in Fig. 11 (c). This method enables end-to-end design of complex amplitude holograms based on \(\mathrm{{THz}}\) metasurfaces, directly mapping the geometric parameters of the unit cells to the holographic image, without the need for full optical modulation. As shown in Fig. 11 (d), the deep neural network module acts as the generator, outputting height and diameter maps for the nanocylinder array corresponding to the target holographic image. These height and diameter maps are then input into the forward physical propagation module, which serves as the discriminator. The propagation module reconstructs the target image on the imaging plane through diffraction propagation. The loss function calculates the difference between the reconstructed image and the target image, driving the training process.

4.3 Sensor

THz metamaterial sensors are highly sensitive devices that utilize the unique electromagnetic responses of \(\mathrm{{THz}}\) metamaterials to detect various chemical substances, biological molecules, or environmental changes [ 92 ]. Traditional design methods for such sensors often rely on human expertise, repeated experimentation, and the identification of the expected functionalities of the THz metamaterial sensors. However, as the performance requirements for these sensors continue to increase, traditional experience-based approaches are no longer sufficient. The trial-and-error process, guided by empirical knowledge, inherently struggles to identify the most representative features from multiple optimization parameters.

In this context, intelligent design emerges as a promising solution to address these limitations. To improve the generalization ability of deep learning models and the design flexibility of THz metamaterial structures, Ge et al.[ 93 ] combined the physical mechanisms of \(\mathrm{{THz}}\) metamaterial sensors with deep learning theory, and proposed a reverse design prediction algorithm for multi-structure \(\mathrm{{THz}}\) metamaterial sensors. They introduced an improved CGAN model for the reverse design of \(\mathrm{{THz}}\) metamaterial sensor structures, as shown in Fig. 12 (a). Additionally, a fully connected forward neural network was proposed to accelerate the verification of the structures generated by the reverse design model. The prediction results using the self-attention conditional wasserstein generative adversarial network (SACW-GAN) model were compared in terms of spectral and image accuracy with those from the standalone CGAN and SAGAN models, in order to validate the effectiveness and accuracy of this method. As shown in Fig. 11 (b), during the training process of the SACW-GAN model, the goal of the discriminator is to learn data features and adjust weights by maximizing the Wasser-stein distance between real and generated samples. The Wasserstein distance measures the disparity between the real and generated samples. Through this process of calculation and analysis, the generator is able to produce pseudo-data that closely matches the real data. The generator and discriminator alternate their training, which enhances the stability of the training process, ensuring that the generated data increasingly resembles the real data. In each iteration of training, the discriminator is updated four times, while the generator is updated once. After the model training is completed, the generated \(\mathrm{{THz}}\) metamaterial sensor structure images are input into the forward neural network for validation, and the absorption curve of the THz metamate-rial structure is calculated. Then, the target absorption curve is compared and analyzed with the generated absorption curve to verify whether the generated \(\mathrm{{THz}}\) metamaterial structure meets the design requirements. The spectral accuracy and image accuracy of the model can reach 95% and \({97}\%\) , respectively, providing significant research value for the inverse design of multi-structure \(\mathrm{{THz}}\) metamaterial sensors.

AI THz sensors have significant research value across various fields. In the design of \(\mathrm{{THz}}\) fingerprint metasurface sensors, Liu et al.[ 94 ] proposed a dual neural network deep learning architecture for the intelligent design of inverted all-dielectric metagrating (IAM), enabling multiplexed sensing of trace fingerprints. As shown in Fig.12 (c), the design of the IAM utilizes silicon material and consists of an inverted grating and waveguide layers. Its performance is influenced by four main parameters: grating period, waveguide layer thickness, grating height, and duty cycle. Through an angle multiplexing mechanism, this design is able to generate high-quality spectral resonances, enhancing the signal of trace analytes. Trace detection is achieved by comparing the spectral changes before and after loading the analyte. The deep learning process for the fingerprint metasensor is shown in Fig. 12 (d). This deep learning design process establishes a customized framework for the fingerprint metasensor by adjusting the coordinate points to generate spectra for specific incident angles. The generated spectra are fitted using interpolation between defined points and serve as the input targets in the deep learning model. As shown in Fig. 12 (e), the system uses an inverse neural network (INN) to predict the metastructure parameters, which are then input into a forward neural network (FNN) for validation. This process effectively addresses the issue of mis-convergence and replaces time-consuming simulations. By employing a “divide and conquer” strategy, the single network is divided into multiple sub-networks, each responsible for prediction in different wavelength ranges. This approach breaks down the larger prediction problem into smaller, more manageable sub-problems, enhancing the overall performance of the network. This method aids in the rapid inverse design and fabrication of IAM sensors based on the given fingerprint peaks of target analytes.

The application of THz metasurfaces in biological sensing and detection is a rapidly advancing area of THz biosensing technology. The key to THz metasurface biosensing technology lies in the use of specially designed THz metasurface structures. By adjusting the structural parameters of the THz metamaterials, the propagation of THz waves can be manipulated, allowing for control over aspects such as focusing, reflection, transmission, and absorption. This facilitates the interaction of \(\mathrm{{THz}}\) waves with biomolecules, enabling high sensitivity and selectivity in the detection of biological samples [ 95 , 96 ]

5 Conclusions

This paper presents a thorough review of key methods in the intelligent design of THz meta-surfaces. These approaches offer significant advantages in terms of physical accuracy and computational efficiency, making them promising for future \(\mathrm{{THz}}\) metasurface and metamate-rial design. We have discussed the role of AI in both forward and inverse design of \(\mathrm{{THz}}\) metasur-faces, highlighting its transformative impact on improving metasurface performance. Specifically, AI-based forward design methods have demonstrated excellent potential, enabling rapid development of \(\mathrm{{THz}}\) metasurface unit cells with high accuracy, and outperforming traditional optical simulation techniques in computational speed. However, despite the significant advancements enabled by AI-driven design methods, several challenges remain.

One of the main research directions for the future is the expansion of AI-based methods to the design of multi-functional \(\mathrm{{THz}}\) metasurfaces. This involves the simultaneous optimization of various optical parameters -such as phase, amplitude, and polarization -to meet the more complex demands of practical applications. As AI technologies continue to improve, these multifunctional metasurfaces are expected to provide more flexible and efficient solutions for a wider range of optical systems.

Another promising area for development is the exploration of real-time adaptive THz meta-surfaces. These metasurfaces could dynamically reconfigure in response to external stimuli or changing operational conditions, offering advantages in applications that require rapid adjustments or responsiveness to environmental variations. Real-time adaptability would enhance the flexibility and functionality of \(\mathrm{{THz}}\) metasurfaces in dynamic environments.

Furthermore, as AI models for inverse design advance, there is growing potential to directly synthesize entire \(\mathrm{{THz}}\) metasurfaces, bypassing the need for pre-designed unit cells. This would provide a more intuitive and efficient design process, allowing for the creation of complex and highly optimized THz metasurfaces. It would streamline the design workflow and open up new possibilities for their applications.

However, despite these advancements, AI-driven design methods face some significant challenges. One of the primary issues is the need for significant computational power. The large-scale training and evaluation of AI models, particularly deep learning models, require substantial computational resources, including high-performance processors like GPUs or TPUs. This can make the use of AI-driven methods prohibitively expensive, especially in environments where computational resources are limited.

Additionally, AI-driven methods tend to consume more energy, especially during the training phase. The computational intensity of AI models leads to higher energy consumption, which could become a limitation for applications where energy efficiency is crucial. Therefore, addressing the high computational and energy demands of AI models should be a major focus for future research, with potential solutions including optimizing models for efficiency, leveraging edge computing, and exploring low-power AI techniques. To move forward, addressing challenges such as model generalization, creating large-scale training datasets, and integrating AI-driven designs with advanced fabrication techniques will be essential. Collaboration between AI, materials science, and nanotechnology will be key to overcoming these barriers, enabling breakthroughs in the rapidly advancing field of THz technology.

References

[1] M. Tonouchi . Cutting-edge terahertz technology [J]. Nature Photon , 2007 . 1 . 97 - 105 .

[2] C. Chaccour , W. Saad , M. Debbah H. V. Poor . Joint sensing, communication, and AI: A trifecta for resilient THz user experiences [J]. IEEE Transactions on Wireless Communications , 2024 . 23 ( 9 ): 11444 - 11460 .

[3] R. Kohlhaas , S. Lauck , M. Deumer , S. Breuer , М. Schell , B. Globisch . Optoelectronic frequency-modulated continuous-wave terahertz spectroscopy with 4THz bandwidth [J]. Nature Communications , 2021 . 12 . 1 - 9 .

[4] Z. Wang , G. Wang , B. Hu , W. Liu , J. Huang , C. Xiong , Y. Zhang , J. Liu , Y. Wang . Fast-printed, large-area and low-cost terahertz metasurface using laser-induced graphene [J]. Carbon , 2022 . 187 . 256 - 265 .

[5] W. Jiang , Q. H. Zhou , J. G. He , M. A. Habibi , S. Melnyk , M. El-Absi . Terahertz communications and sensing for 6G and beyond: A comprehensive review [J]. IEEE Communications Surveys and Tutorials , 2024 . 26 ( 4 ): 2326 - 2381 .

[6] I. F. Akyildiz , C. Han , Z. Hu , S. Nie J. M. Jornet . Terahertz band communication: An old problem revisited and research directions for the next decade [J]. IEEE Transactions on Communications , 2022 . 70 ( 6 ): 4250 - 4285 .

[7] R. Chen , M. Chen , X. Xiao , W. Zhang , J. Li . Multi-user orbital angular momentum based terahertz communications [J]. IEEE Transactions on Wireless Communications , 2023 . 22 ( 9 ): 6283 - 6297 .

[8] H. Zhong , A. Redo-Sanchez , X. C. Zhang . Identification and classification of chemicals using terahertz reflective spectroscopic focal-plane imaging system [J]. Optics Express , 2006 . 14 . 9130 - 9141 .

[9] Y. W. Sun , H. L. Huang , L. X. Shan , S. Fan , Z. Χ. Zhu , X. D. Liu . Quantitative analysis of bisphenol analogue mixtures by terahertz spectroscopy using machine learning method [J]. Food Chemistry , 2021 . 352 . 129313 -

[10] D. D. Ye , W. Z. Wang , H. T. Zhou , H. J. Fang , J. B. Huang , Y. J. Li , H. H. Gong , Z. Li . Characterization of thermal barrier coatings microstructural features using terahertz spectroscopy [J]. Surface and Coating Technology , 2020 . 394 . 125836 -

[11] X. C. Cao , J. H. Hao , Q. Zhao , F. Zhang , J. Q. Fan Z. W. Dong . Analysis of high-frequency atmospheric windows for terahertz transmission along earth-space paths [J]. IEEE Transactions on Antennas and Propagation , 2022 . 70 ( 7 ): 5715 - 5724 .

[12] A. A. Gavdush , F. Kruczkiewicz , B. M. Giuliano , B. Müller , G. A. Komandin , T. Grassi , P. Theulé , K. I. Zaytsev , A. V. Ivlev P. Caselli . Broadband spectroscopy of astrophysical ice analogues [J]. Astronomy and Astrophysics , 2022 . 667 . A49 -

[13] A. B. Haghverdi , A. A. M. Khani , I. Rezaei , T. Aghaee , S. Biabanifard . Graphene ribbons based THz toxic gas sensing [J]. Sensing and Bio-Sensing Research , 2024 . 45 . 100672 -

[14] K. Komatsu , T. Iwamoto , H. Ito , H. Saitoh . THz gas sensing using terahertz time-domain spectroscopy with ceramic architecture [J]. ACS Omega , 2022 . 7 ( 35 ): 30768 - 30772 .

[15] W. Nsengiyumva , S. C. Zhong , L. H. Zheng , W. Liang , B. Wang , Y. Huang . Sensing and non-destructive testing applications of terahertz spectroscopy and imaging systems: State-of-the-art and state-of-the-practice [J]. IEEE Transactions on Instrumentation and Measurement , 2023 . 72 . 1 - 83 .

[16] S. C. Zhong . Progress in terahertz nondestructive testing: A review [J]. Frontiers of Mechanical Engineering , 2019 . 14 . 273 - 281 .

[17] G. Tzydynzhapov , P. Gusikhin , V. Muravev , A. Dremin , Y. Nefyodov , I. Kukushkin . New real-time sub-terahertz security body scanner [J]. Journal of Infrared, Millimeter, and Terahertz Waves , 2020 . 41 . 632 - 641 .

[18] M. Wan , J. J. Healy , J. T. Sheridan . Terahertz phase imaging and biomedical applications [J]. Optics and Laser Technology , 2020 . 122 . 105859 -

[19] A. Gong , Y. Qiu , X. Chen , Z. Zhao , L. Xia , Y. Shao . Biomedical applications of terahertz technology [J]. Applied Spectroscopy Reviews , 2019 . 55 ( 5 ): 418 - 438 .

[20] X. R. Li , J. X. Li , Y. H. Li , A. Ozcan , M. Jarrahi . High-throughput terahertz imaging: progress and challenges [J]. Light Science Application , 2023 . 12 . 233 -

[21] F. Fauquet , F. Galluzzi , P. F. Taday , R. Chapoulie , A. Mounier , A. B. Amara , P. Mounaix . Terahertz time-domain spectro-imaging and hyperspectral imagery to investigate a historical Longwy glazed ceramic [J]. Scientific Reports , 2024 . 14 . 19248 -

[22] J. Hu , S. Bandyopadhyay , Y. H. Liu , L. Y. Shao . A review on metasurface: from principle to smart metadevices [J]. Frontiers in Physics , 2021 . 8 . 586087 -

[23] H.T. Chen , A. J. Taylor , N. Yu . A review of metasurfaces: physics and applications [J]. Reports on Progress in Physics , 2016 . 79 ( 7 ): 076401 -

[24] T. C. Peng , L. M. Si , Q. Q. Yuan , J. Tian , J. X. Deng , T. Y. Ma , X.E. Bao , C. He , W. R. Zhu . Dynamic generation of multiplexed vortex beams by a space-time-coding metasurface [J]. Photonics Research , 2025 . 13 ( 1 ): 225 - 234 .

[25] L. Dong , L. M. Si , B. Y. Liu , Q. T. Shen , R. Niu , X. E. Bao , H. J. Sun , W. R. Zhu . An optically transparent rectifying metasurface for 2.4/5.8 GHz dual-band RF energy harvesting [J]. Applied Physics Letter , 2025 . 126 ( 1 ): 011705 -

[26] A. Li , S. Singh , D. Sievenpiper . Metasurfaces and their applications [J]. Nanophotonics , 2018 . 7 ( 6 ): 989 - 1011 .

[27] X. Jiang , H. Chen , Z. Y. Li , H. K. Yuan , L. Y. Cao , Z. F. Luo , K. Zhang , Z. H. Zhang , Z. G. Wen , L. G. Zhu , X. Zhou , G. F. Liang , D. S. Ruan , L. H. Du , L. F. Wang , G. Chen . All-dielectric metal-ens for terahertz wave imaging [J]. Optics Express , 2018 . 26 . 14132 - 14142 .

[28] Q. Cheng , M. L. Ma , D. Yu , Z. X. Shen , J. Y. Xie , J. C. Wang , N. X. Xu , H. M. Guo , W. Hu , S. M. Wang , T. Li , S. L. Zhuang . Broadband achromatic metalens in terahertz regime [J]. Science Bulletin , 2019 . 64 ( 20 ): 1525 - 1531 .

[29] P. Ding , Y. Li , L. Shao , X. Tian , J.Q. Wang , C.Z. Fan . Graphene aperture-based metalens for dynamic focusing of terahertz waves [J]. Optics Express , 2018 . 26 . 28038 - 28050 .

[30] A. Moreno-Peñarrubia , S. A. Kuznetsov M. Beruete . Ultrathin subterahertz half-wave plate with high conversion efficiency based on zigzag metasurface [J]. IEEE Transactions on Antennas and Propagation , 2020 . 68 ( 11 ): 7700 - 7704 .

[31] Y. Xu , Q. Xu , X. Zhang , X. Feng , Y. Lu , X. Zhang , M. Kang , J. Han , W. Zhang . Stereo metasurfaces for efficient and broadband terahertz polarization conversion [J]. Advanced Functional Materials , 2022 . 32 ( 44 ): 2207269 -

[32] X. L. You , R. T. Ako , W. S. L. Lee , M. Bhaskaran , S. Sriram , C. Fumeaux , W. Withayachumnankul . Broadband terahertz transmissive quarter-wave metasurface [J]. APL Photonics , 2020 . 5 ( 9 ): 096108 -

[33] X. Zhuang , W. Zhang , K. Wang , Y. F. Gu , Y. W. An , X. Q. Zhang , J. Q. Gu . Active terahertz beam steering based on mechanical deformation of liquid crystal elastomer metasurface [J]. Light Science and Applications , 2023 . 12 . 14 -

[34] T. Y. Ma , L. M. Si , C.Y. Dang , R. Niu , G. H. Wu , X. E. Bao , H. J. Sun , W. R. Zhu . Graphene spatiotemporal reconfigurable intelligent surface for terahertz polarization-states manipulation and holographic imaging [J]. Nanoscale Advances , 2025 . 7 ( 7 ): 1825 - 1837 .

[35] S. J. Wang , W. T. Qin , S. Zhang , Y. C. Lou , C. Q. Liu , T. Wu , Q. He , C. S. Tian , L. Zhou , Y. Z. Wu , Z. S. Tao . Nanoengineered spintronic-metasurface terahertz emitters enable beam steering and full polarization control [J]. Nano Letters , 2022 . 22 ( 24 ): 10111 - 10119 .

[36] Y. Ding , X. Chen , Y. Duan , H. Y. Huang , L. D. Zhang , S. Y. Chang , X. X. Guo , X. J. Ni . Metasurface-dressed two-dimensional on-chip waveguide for free-space light field manipulation [J]. ACS Photonics , 2022 . 9 ( 2 ): 398 - 404 .

[37] H. Zhang , L. Dong , P. C. Tang , Q. L. Zhang , H. J. Sun , L. M. Si . Dual-band terahertz chiral metasurface with circularly polarized absorbing and anomalous reflecting [J]. Acta Electronica Sinica , 2023 . 51 ( 10 ): 2690 - 2699 .

[38] Κ. Υ. Guo , M. Zhang , Z. Q. Du , Z. Y. Song . Terahertz spin-selective metasurface for multichannel switching of OAM [J]. Results in Physics , 2024 . 56 . 107279 -

[39] Y. Y. Yuan , K. Zhang , B. Ratni , Q. H. Song , X. Ding , Q. Wu , S. N. Burokur P. Genevet . Independent phase modulation for quadruplex polarization channels enabled by chirality-assisted geometric-phase metasurfaces [J]. Nature Communication , 2020 . 11 . 4186 -

[40] A. U. R. Khalid , F. Feng , N. Ullah , X. Yuan , M. G. Somekh . Exploitation of geometric and propagation phases for spin-dependent rational-multiple complete phase modulation using dielectric metasurfaces [J]. Photonics Research , 2022 . 10 . 877 - 885 .

[41] J. Li , Y. Yuan , Q. Wu K. Zhang . Bi-isotropic huygens' metasurface for polarization-insensitive cross-polarization conversion and wavefront manipulation [J]. IEEE Transactions on Antennas and Propagation , 2024 . 72 ( 3 ): 2445 - 2454 .

[42] W. Song , X. Liang , S.Q. Li , D.D. Li , R. Paniagua-Domínguez , K. H. Lai , Q.Y. Lin , Y.J. Zheng , A. I. Kuznetsov . Large-scale huygens' metasurfaces for holographic 3D near-eye displays [J]. Laser and Photonics Reviews , 2021 . 15 ( 9 ): 2000538 -

[43] Z. C. Liu , D. Y. Zhu , S. P. Rodrigues , K. T. Lee , W. Cai . Generative model for the inverse design of metasurfaces [J]. Nano Letters , 2018 . 18 ( 10 ): 6570 - 6576 .

[44] H. M. Nemat-Abad , E. Zareian-Jahromi , R. Basiri . Design of metasurface-based multi-layer THz filters utilizing optimization algorithm with distinct fitness function definitions [J]. Plasmonics , 2021 . 16 . 1865 - 1876 .

[45] Y. Teng , C. Li , S. C. Li , Y. H. Xiao , L. Jiang . Efficient design method for terahertz broadband metasurface patterns via deep learning [J]. Optics and Laser Technology , 2023 . 160 . 109058 -

[46] Z. Deng , Y. Li , Y. Li , Y. Wang , W. Li , Z. Zhu , C. Guan , J. Shi . Diverse ranking metamaterial inverse design based on contrastive and transfer learning [J]. Optics Express , 2023 . 31 ( 22 ): 32865 - 32874 .

[47] L. Li , H. Ruan , C. Liu , Y. Li , Y. Shuang , A. Alù , C. W. Qiu , T. J. Cui . Machine-learning reprogrammable metasurface imager [J]. Nature Communications , 2019 . 10 . 1802 -

[48] S. Nishijima Y. Monnai . Topology-optimized metal structure for broadband wavefront engineering of guided terahertz waves [J]. IEEE Transactions on Terahertz Science and Technology , 2024 . 14 ( 4 ): 495 - 501 .

[49] W. Y. Ji , J. Chang , H. X. Xu , J. R. Gao , S. Groblacher , H. P. Urbach , A. J. L. Adam . Recent advances in metasurface design and quantum optics applications with machine learning, physics-informed neural networks, and topology optimization methods [J]. Light: Science and Applications , 2023 . 12 . 169 -

[50] Z. Li , R. Pestourie , Z. Lin , S. G. Johnson , F. Capasso . Empowering metasurfaces with inverse design: Principles and applications [J]. ACS Photonics , 2022 . 9 ( 7 ): 2178 - 2192 .

[51] R. C. Zhu , T. S. Qiu , J. F. Wang , S. Sui , C. L. Hao , T. H. Liu , Y. F. Li , M. D. Feng , A. Zhang , C. W. Qiu , S. B. Qu . Phase-to-pattern inverse design paradigm for fast realization of functional metasurfaces via transfer learning [J]. Nature Communications , 2021 . 12 . 2974 -

[52] W. J. Zhou , S. Y. Wang , Q. Wu , X. C. Xu , X. J. Huang , G. L. Huang , Y. Liu , Z. Fan . An inverse design paradigm of multi-functional elastic metasurface via data-driven machine learning [J]. Materials and Design , 2023 . 226 . 111560 -

[53] J. Peurifoy , Y. Shen , L. Jing , Y. Yang , F. Cano-Renteria , B. G. DeLacy , J. D. Joannopoulos , M. Tegmark , M. Soljačić . Nanophotonic particle simulation and inverse design using artificial neural networks [J]. Science Advance , 2018 . 4 ( 6 ): 4206 -

[54] Μ. Μ. Elsawy , S. Lanteri , R. Duvigneau , J. А. Fan , P. Genevet . Numerical optimization methods for metasurfaces [J]. Laser Photonics Reviews , 2020 . 14 ( 10 ): 1900445 -

[55] S. Katoch , S. S. Chauhan , V. Kumar . A review on genetic algorithm: past, present, and future [J]. Multimedia Tools and Applications , 2021 . 80 . 8091 - 8126 .

[56] Y. Wang , G. Wu , J. Zhang , X. Wu , G. Yuan , J. Liu . Genetic algorithm-enhanced design of ultra-broadband tunable terahertz metasurface absorber [J]. Optics and Laser Technology , 2024 . 170 . 110262 -

[57] S. Jafar-Zanjani , S.Inampudi , H. Mosallaei . Adaptive genetic algorithm for optical metasurfaces design [J]. Scientific Reports , 2018 . 8 . 11040 -

[58] K. Deb H. Jain . An evolutionary many-objective optimization algorithm using reference-point-based nondominated sorting approach, part I: Solving problems with box constraints [J]. IEEE Transactions on Evolutionary Computation , 2014 . 18 ( 4 ): 577 - 601 .

[59] A. Tharwat W. Schenck . A conceptual and practical comparison of PSO-style optimization algorithms [J]. Expert Systems with Applications , 2021 . 167 . 114430 -

[60] J. R. Thompson , H. D. Nelson-Quillin , E. J. Coyle , J. P. Vernon , E. S. Harper , M. S. Mills . Particle swarm optimization of polymer-embedded broadband metasurface reflectors [J]. Optics Express , 2021 . 29 ( 26 ): 43421 - 43434 .

[61] W. T Xing , L. M. Si , L. Dong , H. Zhang , T. Y. Ma , H. J. Sun , X. Bao , J. Ding . Rapid design of hybrid mechanism metasurface with random coding for terahertz dual-band RCS reduction [J]. Optics Express , 2023 . 31 . 28444 - 28458 .

[62] L. Ma , Z. H. Wang , L. H. Feng , W. Dong , W. L. Guo . Constrained multi-objective optimization problem model to design multi-band terahertz metamaterial absorbers [J]. Optical Materials Express , 2023 . 13 . 739 - 751 .

[63] Η. Τ. Kollmann , D. W. Abueidda , S. Koric , E. Guleryuz , N. A. Sobh . Deep learning for topology optimization of 2D metamaterials [J]. Materials and Design , 2020 . 196 . 109098 -

[64] J. L. Cui , B. Z. Zhang , H. C. Xu , H. X. Sun , G. R. Shao . Topological flexible metamaterials with isotropic dual-frequency terahertz-band absorption [J]. Optics Communications , 2019 . 441 . 90 - 95 .

[65] K. Rouhi , A. Abdolali , S. Fallah . Designing approach of terahertz broadband backscattering reduction based on combination of diffusion and absorption [J]. Optik , 2021 . 246 . 167771 -

[66] S. So , T. Badloe , J. Noh , J.B. Abad , J. Rho . Deep learning enabled inverse design in nanophotonics [J]. Nanophotonics , 2020 . 9 ( 5 ): 1041 - 1057 .

[67] W. Ma , Z. C. Liu , Z. A. Kudyshev , A. Boltasseva , W. S. Cai , Y. M. Liu . Deep learning for the design of photonic structures [J]. Nature Photonics , 2021 . 15 . 77 - 90 .

[68] L. M. Si , R. Niu , C. Y. Dang , X. E. Bao , Y. Q. Zhuang , W. R. Zhu . Advances in artificial intelligence for artificial metamaterials [J]. APL Materials , 2024 . 12 ( 12 ): 120602 -

[69] Y. L. Cun , Y. Bengio , G. Hinton . Deep learning [J]. Nature , 2015 . 521 . 436 - 444 .

[70] C. Janiesch , P. Zschech , K. Heinrich . Machine learning and deep learning [J]. Electronic Markets , 2021 . 31 . 685 - 695 .

[71] C. C. Nadell , B. H. Huang , J. M. Malof , W. J. Padilla . Deep learning for accelerated all-dielectric metasurface design [J]. Optics Express , 2019 . 27 . 27523 - 27535 .

[72] R. Singh , E. Plum , C. Menzel , C. Rockstuhl , A. Κ. Azad , R. A. Cheville , F. Lederer , W. Zhang , N. I. Zheludev . Terahertz metamaterial with asymmetric transmission [J]. Physical Review B , 2009 . 80 . 153104 -

[73] F. Gao , Z. Zhang , Y. F. Xu , L. Y. Zhang , R. Q. Yan , X. F. Chen . Deep-learning-assisted designing chiral terahertz metamaterials with asymmetric transmission properties [J]. Journal of the Optical Society of America B , 2022 . 39 ( 6 ): 1511 - 1519 .

[74] J. Cheng , R. Li , Y. Wang , Y. Yuan , X. Wang , S. Chang . Inverse design of generic metasurfaces for multifunctional wavefront shaping based on deep neural networks [J]. Optics and Laser Technology , 2023 . 159 . 109038 -

[75] Y. Li , M. Deng , Z. Fang , P. Peng , Y. Chen , Z. Fang . Inverse design of unidirectional transmission nanostructures based on unsupervised machine learning [J]. Advanced Optical Materials , 2022 . 10 ( 12 ): 2200127 -

[76] Y. Wang , Y. Zhu , Z. Cui , H. Jiang , K. Zhang , X. Wang . Ensemble learning: a bidirectional framework for designing data-driven THz composite metamaterials [J]. Journal of the Optical Society of America B , 2022 . 39 . 835 -

[77] Z. Liu , D. Zhu , S. P. Rodrigues , K. T. Lee , W. Cai . A generative model for inverse design of metamaterials [J]. Nano letter , 2019 . 18 ( 10 ): 6570 - 6576 .

[78] W. B. He , M. Y. Tong , Z. J. Xu , Y. Hu , Χ. Α. Cheng , T. Jiang . Ultrafast all-optical terahertz modulation based on an inverse-designed metasurface [J]. Photonics Research , 2021 . 9 . 1099 - 1108 .

[79] Z. Hou , C. L. Zheng , J. Li , P. Y. Zhang , S. Z. Li , S. P. Zheng , J. Shen , J. Q. Yao , C. Y. Li . On-demand design based on deep learning and phase manipulation of all-silicon terahertz chiral metasurfaces [J]. Results in Physics , 2022 . 42 . 106024 -

[80] S. Liu P. Acar . Generative adversarial networks for inverse design of two-dimensional spinodoid metamaterials [J]. AIAA Journal , 2024 . 62 ( 7 ): 1 - 10 .

[81] Β. Χ. Mao , H. H. Chang , X. H. Xing , Q. K. Zhang , D. Zou , Y. Liu , J. Q. Yao , H. X. Bi , L. Wu . Designing ultra-broadband terahertz polarization converters based on the transformer model [J]. Optics Communications , 2024 . 559 ( 15 ): 130434 -

[82] J. T. Lv , R. Z. Zhang , C. L. Liu , Z. M. Ge , Q. C. Gu , F. Feng , G. Y. Si . Polarization-controlled metasurface design based on deep ResNet [J]. Optics and Laser Technology , 2024 . 179 . 111396 -

[83] S. Chen , Z. Y. Hou , J. J. Wang , Y. H. Qiu , Z. Q. Li , J. Shen . Chiral metasurface vortex focusing in terahertz band based on deep learning [J]. IEEE Photonics Technology Letters , 2023 . 35 ( 11 ): 637 - 640 .

[84] J. Q. Luan , Y. J. Zhang , Y. Chen , D. S. Gao , P. L. Li , J. Q. Li . Genetic algorithm based terahertz multifunctional reconfigurable Dirac semimetallic coded metasurface [J]. Acta Physica Sinica , 2024 . 73 . 144204 -

[85] P. Peng , Z. C. Liu , C. G. Dai , X. He , D. Y. Wang , Z. B. Dang , Y. X. Chen , Y. C. Dai , H. Zhang , S. L. Sun , L. Zhou , Z. Y. Fang . Arbitrary multifunctional vortex beam designed by deep neural network [J]. Advanced Optical Materials , 2024 . 12 ( 6 ): 2300158 -

[86] Y. J. Tan , C. Y. Zhu , T. C. Tan , A. Kumar , L. J. Wong , Y. Chong , R. Singh . Self-adaptive deep reinforcement learning for THz beamforming with silicon metasurfaces in 6G communications [J]. Optics Express , 2022 . 30 . 27763 - 27779 .

[87] R. Zhao , L. Huang , Y. Wang . Recent advances in multi-dimensional metasurfaces holographic technologies [J]. PhotoniX , 2020 . 1 . 20 -

[88] A. Márquez , C. Li , A. Beléndez , S. A. Maier , H. Ren . Information multiplexing from optical holography to multi-channel metaholography [J]. Nanophotonics , 2023 . 12 ( 24 ): 4415 - 4440 .

[89] M. Yamaguchi , H. Saito , S. Ikezawa , K. Iwami . Highly-efficient full-color holographic movie based on silicon nitride metasurface [J]. Nanophotonics , 2024 . 13 ( 8 ): 1425 - 1433 .

[90] Z. Y. Hou , P. Y. Zhang , S. X. Chen , J. J. Wang , Y. H. Qiu , T. T. Tang , C. Y. Li . On-demand design of holographic metasurfaces and continuous phase and amplitude modulation method based on deep learning [J]. Results in Physics , 2024 . 66 . 108026 -

[91] W. Wei , P. Tang , J. Z. Shao , J. Zhu , X. Y. Zhao , C. Z. Wu . End-to-end design of metasurface-based complex-amplitude holograms by physics-driven deep neural networks [J]. Nanophotonics , 2022 . 11 ( 12 ): 2921 - 2929 .

[92] W. D. Xu , L. J. Xie , Y. B. Ying . Mechanisms and applications of terahertz metamaterial sensing: a review [J]. Nanoscale , 2017 . 37 ( 9 ): 13864 - 13878 .

[93] H. Ge , Y. Bu , X. Ji , Y. Jiang , K. Jia , Y. Zhang , Y. Zhang , X. Wu , Q. C. Sun . Inverse design of multistructured terahertz metamaterial sensors based on improved conditional generative network [J]. ACS Applied Materials Interfaces , 2024 . 16 ( 44 ): 60772 - 60782 .

[94] X. Liu , Y. Xie , Y. Yan , Q. Niu , L. G. Zhu , Z. Dong , Q. H. Liu , J. Zhu . Rapid on-demand design of inverted all-dielectric metagratings for trace terahertz molecular fingerprint sensing by deep learning [J]. ACS Photonics , 2024 . 11 ( 11 ): 4838 - 4845 .

[95] X. Qiao , A. Wang , W. Xiao , J. Zhang , X. Xia , Y. Liang , P. Jiang . A simple structured and homo-frequency point mode switchable THz full-space metasurface based on temperature-controlled vanadium dioxide [J]. Optics and Laser Technology , 2024 . 170 . 110233 -

[96] Y. Hu , X. Wang , X. Yang , H. Duan . All-dielectric metasurfaces for polarization manipulation: principles and emerging applications [J]. Nanophotonics , 2020 . 9 ( 13 ): 3755 - 3780 .

Funding

National Key R and D Program of China(2022YFF0604801)

National Natural Science Foundation of China(62271056)

National Natural Science Foundation of China(62171186)

National Natural Science Foundation of China(62201037)

Technology Innovation Center of Infrared Remote Sensing Metrology Technology of State Administration for Market Regulation(AKYKF2423)

Beijing Natural Science Foundation of China-Haidian Original Innovation Joint Fund(L222042)

Open Research Fund of State Key Laboratory of Millimeter Waves(K202326)

111 Project of China(B14010)