With the deepening understanding of the visible world, human interest in the invisible realm has become increasingly strong. Electromagnetic waves have endowed us with remarkable perceptual capabilities, such as improving our ability to detect distant objects through meter wave and millimeter wave radars [1], achieving long-range communication through short-wave technology [2], and enabling us to see in the dark through infrared imaging [3,4]. X-rays have provided unprecedented insight into our internal structures [5], while the advent of spatial target sensing [6–8] has further expanded our understanding of electromagnetic signals. In recent years, significant progress has been made in wireless perception technologies, yielding promising outcomes such as wireless charging [9–11], 6G [12,13], and WIFI7 [14] communications, and star chain systems. However, the rapid, precise, and parallel implementation of spatial location sensing remains one of the most challenging hurdles in this field. Traditional direction of arrival (DOA) estimation algorithms, including spatial spectral estimation [15], maximum entropy spectral estimation, and high-resolution spectral analysis [16], typically rely on large antenna arrays, resulting in high construction and maintenance costs. Although methods like multiple signal classification (MUSIC) [17] and rotational invariance estimation (ESPRIT) [18] have been proposed, they often demand substantial computational resources or suffer from real-time performance degradation due to extensive computations. Moreover, the number of detectable targets is constrained by the limited number of antenna array elements, which tremendously restricts the effectiveness of this technology.

Metasurfaces have garnered significant attention in the realms of invisibility cloaks [19,20], perfect absorbers [21,22], holograms [23,24], metasurface lenses [25,26], wireless communications [27,28], and electromagnetic sensing [29–32], which is attributed to their remarkable capacity for manipulating electromagnetic waves. Some researchers have leveraged metasurfaces to enable communication signal demand sensing and management in smart cities [33]. Additionally, some efforts are devoted to achieving direction-specific electromagnetic sensing based on a human gaze utilizing metasurfaces [34]. Furthermore, researchers have demonstrated direction of arrival (DOA) sensing of one or two targets in half-space [35–37] using programmable metasurfaces [38–44] as electromagnetic receiving mediums. However, based on the literature reviewed by the authors, existing schemes are difficult to simultaneously and accurately perceive multiple targets in full space, presenting a pressing challenge given the substantial number of targets required in practical applications. Moreover, many existing studies have predominantly focused on single-function design paradigms, thus neglecting to efficiently harness sensing information. The above limitations significantly diminish the practicality and systematicity of the research endeavors.

In this paper, we present a mirror metasurface-based target sensing and beamforming system. The system comprises three modules: signal receiver module, target sensing module, and beamforming module. In the signal receiver module, we employ a nine-beam mirror metasurface, a reverse-engineered component of a genetic algorithm (GA), as the electromagnetic signal receiving device to collect and integrate electromagnetic signals emitted or reflected by space targets. The target sensing module utilizes probes to measure the near-field electromagnetic response of a metasurface at multiple points and generate a combined real-imaginary data sequence of electric field strengths corresponding to the number and location of targets. In the processing program, the data sequence undergoes input into an intelligent resolution framework based on a deep neural network combined with a genetic algorithm (DNN-GA). This framework outputs spatial location sensing results of the targets based on the fingerprint spectra that correspond to the number, location, and electromagnetic response of each target. The beamforming module involves the rapid optimization and design of topological patterns using GA according to perceived target locations and practical requirements. These patterns may consist of tracking targets mode (mode T), avoiding targets mode (mode A), hybrid mode combining tracking and avoidance (mode H), or reducing RCS mode (mode R) (see Fig. 1).

Compared to existing work, our contributions mainly include two aspects. First, our system achieves position perception for any number of targets in full space, enhancing spatial perception capabilities. Second, we integrate target sensing and beamforming into a comprehensive system, which makes the proposed system possess higher practicality. As a result, our designed system may reduce costs and time associated with sensing numerous targets in space, which can be applied in various modern and complex scenarios such as indoor smart wireless charging, smart driving sensing, anti-trunking operations, and smart city construction.

GA is utilized in the reverse design of the array to generate a binary phase profile of a metasurface that combines sensing range and accuracy. The design process consists of dividing the space, designing mirror beams, and optimizing a far-field pattern.

First, to achieve full-space perception, we employ the polarization separation function of a mirror metasurface for spatial segmentation. We can receive orthogonal polarized waves of the incident electromagnetic wave in the transmitted space, so we can determine the target’s half-space location based on the polarization of the received electromagnetic wave. Second, we use GA to design mirror beams and take into account both transmission and reflection when deriving directional pattern expressions, so that the optimized beam is completely mirror-symmetric in both transmission and reflection half-spaces, significantly reducing optimization complexity and conserving time and computational resources. The mirrored beam simplifies subsequent algorithms by requiring only a half-space recognition algorithm, combined with previous spatial segmentation results, to achieve full-space perception. Third, evaluation indices of sensing range and accuracy are introduced to optimize a far-field pattern, so that the designed metasurface simultaneously possesses the broader sensing range and higher accuracy. Due to conflicting energy distribution for the two metrics, we must seek a balance point.

Ultimately, the nine-beam mirrored metasurface emerges as the optimal choice. The main beam enhances angular resolution compared to non-directional uniform energy distribution, while multiple beams offer greater half-space coverage, facilitating target perception in space. Given the mirroring property of the metasurface, we discuss only the case of incidence in the upper half-space followed by reception in the lower half-space, and the principle and algorithm are completely the same in the full space. The schematic Fig. 2(a) illustrates the structure of a 1-bit polarized rotating unit cell with a unit cell period $p=15\mathrm{mm}$. The top and bottom layers consist of mutually perpendicular identical polarized grids with metal strips of width $l_{\mathrm{top}}=2.8\mathrm{mm}$ and a gap width $d=2.8\mathrm{mm}$. The middle layer features an Egyptian-axe pattern with a line width $s=1\mathrm{mm}$ and an opening width $g=1\mathrm{mm}$, and is connected in the middle by metal strips with a width $l_{\mathrm{mid}}=2\mathrm{mm}$. The three metal layers are separated by two layers of dielectric material F4B with a thickness $h=2\mathrm{mm}$. This structure achieves a 180° phase difference by rotating at an angle $\theta =\pm 45°$. Simulated transmission phase and transmission amplitude curves [see Fig. 2(b)] of the unit cell at 8–12 GHz demonstrate stable maintenance of a 180° phase difference within this frequency band, with cross-polarized transmittance consistently above 94%. These results provide a foundation for the formation of an efficient transmission array. Figure 2(d) illustrates the process of optimizing the phase distribution of the array through the GA. It indicates that the optimal individual and average individual scores of the population converge at 30 generations. The optimization search spans 90 generations to achieve a more optimal phase distribution. The phase topology diagram of an array composed of 1-bit rotation units is optimized using GA, and it produces nine beams in the expected direction [see Figs. 2(e)–2(h)].

We use probes to receive the transmitted electric field intensity and transmit it to a computer, forming a fingerprint spectrum that correlates one-to-one with position information. The fingerprint spectrum data is processed by GA enhanced by DNN, for processing the fingerprint spectrum and intelligently recognizing the number and positions of targets. Specifically, after the electromagnetic waves emitted (active) or reflected (passive) by targets in space are captured by metasurfaces, the difference in position information is amplified and a new transmitted beam is generated. The probe placed in the transmitted space can easily obtain the characteristic information of the transmitted beam. This feature information presented as a sequence of electric field strengths is stored in a computer for subsequent analysis and processing in subsequent modules. Initially, we adhere to the theorem of superposition of electric field strength vectors, as illustrated in Eq. (1), where the binary coefficient $k_i$ indicates the presence of a target in the integration region, and ${\mathit{E}}_i$ represents the electric field information received by the probe when the target exists independently. In essence, the electric field derived from sampling is expanded in the base space comprising individual responses of the electric field at each angle, where $k_i$ serves as the corresponding expansion coefficient: $${\mathit{E}}_{\mathrm{prob}}(X,Y)={\int }_0^{2\pi }{\int }_0^{\frac{\pi }{2}}{k}_i{\mathit{E}}_i{|}_{({\theta }_i,{\phi }_i)}\mathrm{d}\theta \mathrm{d}\phi .$$(1)

Based on this approach, we have successfully addressed the direct determination of target number and position via the target sensing module. However, the number of spatial bases reaches 360 × 90 when the recognition accuracy is 1°, which will seriously reduce the convergence speed of GA and the accuracy of optimization results. To overcome this hurdle, we introduce DNN to preprocess the sampled electric field data, and predict the number of 1 and 0 in the basis coefficients, which substantially enhances the performance of GA. As illustrated in Fig. 3(a), the incorporation of DNN-GA results in a remarkable increase in target perception accuracy by over 40% compared to the conventional GA. Clearly, the addition of DNN not only fortifies GA’s capabilities, but also expedites the process, laying the groundwork for the realization of the high-accuracy target sensing module. As shown in Fig. 3(f), 10 targets are randomly generated for numerical simulation verification (marked in magenta), with the results of target position sensing generally aligning with expectations.

For the beamforming module, we implement an improved GA framework to devise the metasurface array, comprising an arbitrary bit programmable metasurface to achieve four modes: mode T, mode A, mode H, and mode R. In practical scenarios, beamforming typically aligns with mission requirements following target position sensing. The GA framework employed in signal receiver module for designing the nine-beam array remains applicable in this module after slight modifications to the fitness function: $$f_{\mathrm{adapt}}^1=-\sum _{i=1}^n{|{\mathit{E}}_i|}_{({\theta }_i,{\phi }_i)}|,$$(2)$$f_{\mathrm{adapt}}^2=\sum _{i=1}^n{|{\mathit{E}}_i|}_{({\theta }_i,{\phi }_i)}|,$$(3)$$f_{\mathrm{adapt}}^3=-\sum _{i=1}^l{|{\mathit{E}}_i|}_{({\theta }_i,{\phi }_i)}|+\sum _{j=1}^m{|{\mathit{E}}_j|}_{({\theta }_j,{\phi }_j)}|,$$(4)$$f_{\mathrm{adapt}}^4=-{\int }_{\theta =0}^{180}{\int }_{\phi =0}^{360}{|{\mathit{E}}_i|}_{({\theta }_i,{\phi }_i)}|\mathrm{d}\theta \mathrm{d}\phi .$$(5)

Equations (2)–(5) delineate the mode T, mode A, mode H, and mode R, respectively, with $|{\mathit{E}}_i|$ denoting the modal value of electric field strength in the given direction.

We validate the simulation of the signal receiver module, target sensing module, and beamforming module individually on a computer platform. Initially, to ascertain the validity of the signal receiver module and target sensing module, we conduct full-wave EM simulations using CST and process the results using MATLAB. By establishing 100 groups of feeders with random numbers and spatial locations as unknown targets, the electric field strength at the sampling position after the electromagnetic wave passes through the metasurface is calculated. The sampling range encompasses the line between ($-150$, 0, $-10$) and (150, 0, $-10$). Considering the polarization-rotation property of the metasurface, the sampled electric field is the orthogonal component of the incident electric field. The resultant sampling data is fed into the trained DNN, yielding an estimated number of targets. To verify the improvement in electromagnetic sensing accuracy of the GA guided by DNN proposed in this section, we randomly set 100 sets of 10-target sensing scenarios and obtain their electromagnetic response vectors through numerical simulation. On the one hand, the GA is directly used to solve the expansion coefficients of the spatial basis function expansion model, and on the other hand, the GA guided by DNN is used, that is, the target number is first perceived through DNN, and then the target number and electromagnetic response vector are input into the spatial basis function expansion model to solve the expansion coefficients. The comparison of the solution accuracy is shown in Fig. 3(d), where it can be clearly seen that after introducing DNN, the GA’s perception accuracy of the target location has been greatly improved, which once again verifies the effectiveness of the DNN-GA target perception framework designed in this section.

Second, we utilize the simplest patch as a 2-bit unit cell for beamforming module validation and employ the GA framework for designing the metasurface array on the MATLAB platform. We assume the presence of three targets in space located at (27°, 0°), (10°, 137°), and ($-23°$, 42°). By integrating the fitness function defined in Eqs. (2)–(5) into the GA framework, we optimize the phase distribution of the metasurface array. The simulation results depicted in Fig. 4 are in coincidence with expectations, which verifies the high efficiency and accuracy of the beamforming module.

To further validate the system, we establish the test setup in the microwave anechoic chamber, as illustrated in Fig. 5(a). Due to equipment constraints, we can only visualize the detection results of single and dual targets. A microwave horn placed in any position serves as the unknown target, while a metasurface, fabricated using conventional printed circuit board technology, functions as the signal receiving module. Data from the vector network analyzer is fed to a connected computer, where the DNN-GA framework processes the output data and visualizes the number and location of targets. The resulting visualization, depicted in Fig. 5(c), indicates a slight reduction in recognition accuracy compared to the simulation. This discrepancy may be attributed to manufacturing errors, electromagnetic interference in space, and signal sampling errors. Nonetheless, the test results affirm the effectiveness of the modules and convincingly validate the practicality of the system.

It is worth noting that since the samples of the signal receiver module are fabricated and measured to prove the effectiveness of the GA framework, the beamforming module that has been validated in simulation will no longer undergo repeated testing.

In conclusion, we present a fast and precise target sensing and beamforming system based on the enhanced GA, the DNN, and the metasurface with mirror beams. This system comprises three modules: signal receiver module, target sensing module, and beamforming module. In the signal receiver module, a metasurface has been meticulously optimized by the GA to enhance the range and accuracy of spatial angle recognition. The target sensing module measures the electric field intensity transmitted from the metasurface, and the sampling information captured by the DNN-GA serves as the unique identification ID of the target position. In the beamforming module, an improved GA framework combined with 2-bit unit cells is demonstrated, and the metasurface realizes four modes based on actual requirements.

The simulation and test results demonstrate that the proposed target sensing and beamforming system has great practical application value. Compared to traditional DOA detection systems, our design scheme improves recognition speed and the ability to identify the number of targets, while reducing system complexity and cost of manufacturing and maintenance. In comparison to traditional beamforming systems, our design scheme with four flexible working modes can adapt to a wider range of practical needs, which has bright application prospects in modern wireless communication, wireless sensing, and smart cities.