A machine learning based terahertz graphene tunable metasurface design method

Abstract

The application discloses a machine learning-based terahertz graphene adjustable super surface design method, which comprises the following steps: setting a first target matrix and a second target matrix for phase optimization according to target intensity and position of a far-field diffraction beam; constructing a phase optimization model based on the combination of simulated annealing and Fourier transform, using the first target matrix to obtain optimal phase distribution of the super surface through iterative optimization; constructing an amplitude-phase joint optimization model based on particle swarm optimization, using the second target matrix to jointly optimize the amplitude and phase of the super surface, obtaining a corrected disturbance phase distribution and a quantized amplitude distribution, and realizing dynamic switching of the number of far-field beams; the machine learning algorithm is used as the basis, the optimization model has universality, and the graphene super surface unit has a relatively deep modulation depth, so that the number of far-field diffraction beams can be arbitrarily changed and switched.

Description

Technical Field

[0001] This invention relates to the field of metasurface design technology, and more specifically to a terahertz graphene tunable metasurface design method based on machine learning. Background Technology

[0002] Terahertz (THz) waves, with a frequency range from 0.1 to 10 THz, are known as the "terahertz gap." With the rapid development of terahertz technology, these waves have shown great potential in areas such as precise imaging, non-destructive detection, and long-range sensing. However, due to the unique characteristics of the terahertz frequency band, existing technologies still face many challenges in beam manipulation, particularly in multi-beam manipulation and beam number switching.

[0003] Existing beam-controlled metasurface design methods primarily rely on manual encoding and array arrangement simulation using electromagnetic software. However, with increasing demands for far-field control functionality, this approach proves inadequate, failing to provide precise beam control. More importantly, compared to phase control, there is less research on amplitude control. The impact of amplitude on far-field diffraction needs further investigation. In complex applications with multiple beam count switching, joint optimization of both is essential. Existing phase and amplitude recovery algorithms typically require tedious iterative calculations and struggle to guarantee stable performance and optimization results in variable material structures.

[0004] Furthermore, the design of terahertz beam control devices primarily relies on metasurfaces made of traditional materials. These metasurfaces control beam separation through phase modulation, but due to limitations in structural size and geometry, they cannot independently control amplitude and phase, making it difficult to meet the requirements for efficient beam control. In recent years, tunable materials, such as graphene, photosensitive silicon, and liquid crystals, have begun to be incorporated into the design of metasurface unit structures. However, their modulation characteristics are difficult to balance. In far-field multi-beam control, general unit structures often cannot achieve efficient and accurate beam separation and quantity switching in large pixel arrays.

[0005] In summary, current terahertz beam control devices mainly suffer from problems such as the inability to switch the number of beams, poor uniformity of the main beam, inability to adjust device parameters, and redundant and complex calculation methods. Summary of the Invention

[0006] Purpose of the invention: The purpose of this invention is to provide a terahertz graphene tunable metasurface design method based on machine learning, to realize a beam splitter with adjustable number of beams in a wide bandwidth, to demonstrate arbitrary tunability of beams and uniformity among multiple beams, to provide a general multi-beam switchable metasurface design method, and to solve the problems existing in the background technology.

[0007] Technical solution: The present invention provides a terahertz graphene tunable metasurface design method based on machine learning, comprising the following steps:

[0008] (1) Based on the target intensity and position of the far-field diffracted beam, set the first target matrix and the second target matrix for phase optimization; wherein, the second target matrix is ​​used for amplitude and phase joint optimization;

[0009] (2) Construct a phase optimization model based on the combination of simulated annealing and Fourier transform, and use the first objective matrix to obtain the optimal phase distribution of the metasurface through iterative optimization;

[0010] (3) Construct a joint amplitude and phase optimization model based on particle swarm optimization, and use the second objective matrix to jointly optimize the amplitude and phase of the metasurface to obtain the corrected perturbation phase distribution and quantized amplitude distribution, thereby realizing the dynamic switching of the number of far-field beams.

[0011] Furthermore, in step (2), the phase optimization model uses the simulated annealing algorithm as a framework, combines Fourier transform to perform far-field calculation and phase update, and gradually approximates the optimal phase distribution corresponding to the target far field through random perturbation, temperature decay and error acceptance criteria.

[0012] Furthermore, in step (3), the amplitude-phase joint optimization model takes the particle swarm optimization algorithm as its core, uses amplitude and phase as joint optimization variables, minimizes the error between the theoretical far field and the target far field through global search, and performs multi-level quantization on the amplitude after optimization to generate a discrete amplitude set that matches the control capability of the metasurface unit.

[0013] Furthermore, the method is universal; the optimized amplitude and phase distribution can be adapted to various tunable materials, and by customizing the quantization order and threshold, it can meet the requirements of different hardware platforms and control precision.

[0014] The terahertz multi-beam tunable metasurface described in this invention is implemented using any of the methods described above. The metasurface is composed of multiple periodically arranged graphene metasurface units. Each unit includes a bottom metal reflector, an intermediate flexible dielectric layer, and a top open-ring pattern composed of metal and graphene. Phase modulation is achieved by rotating the angle of the open-ring pattern, and amplitude modulation is achieved by adjusting the Fermi level of the graphene material. Amplitude and phase modulation are independent of each other.

[0015] Furthermore, the metasurface unit has the ability to modulate amplitude with high modulation depth over a wide frequency band. By changing the Fermi level, the reflection amplitude can be continuously adjusted over a wide range while maintaining a stable phase response. By rotating the open ring pattern, phase coverage of 0°-360° can be achieved, meeting the requirements of multi-beam dynamic control.

[0016] Furthermore, the metasurfaces are arranged in an array according to the optimal phase distribution and quantized amplitude distribution output by the optimization model. By setting the Fermi level and the rotation angle of the open ring of each graphene unit, the number, position and intensity of the far-field diffraction beams can be arbitrarily and dynamically switched.

[0017] An electronic device according to the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when loaded onto the processor, implements any of the methods described herein.

[0018] The present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements any of the methods described herein.

[0019] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) The present invention proposes a JAP-PSO optimization model based on machine learning, which realizes the joint optimization of amplitude and phase through particle swarm optimization, and combines quantization processing to optimize the amplitude index and threshold, minimizes the mean square error (MSE) between the far-field target matrix and the actual matrix, and performs far-field calculation and result visualization at the same time. The advantage of JAP-PSO is that it can effectively avoid local optimal solutions by utilizing the global search capability of the algorithm, and ensures the directivity and uniformity of the far-field beam through multiple optimizations and joint optimization of amplitude and phase. The optimized far-field calculation and visualization intuitively analyze the optimization effect, ensuring the accuracy of the target far-field and the calculation results. (2) The JAP-PSO optimization model extracts optical physical properties such as amplitude and phase from the material, that is, the optimization result of the algorithm can be applied to different materials, such as tunable materials such as graphene, liquid crystal, and germanium. The multi-order discrete quantization process in the optimization can be adapted to different materials. The optical properties are customized, so the model can be widely applied to optical beam control, wireless communication, antenna design and other fields, and can provide efficient and flexible solutions, and adapt to different precision requirements and hardware constraints. (3) The metasurface unit designed in this invention is composed of metal and graphene in the top open ring pattern. In terms of phase, the maximum phase difference is no more than 20° by rotating the angle of the pattern, which can completely cover a period of 360°; in terms of amplitude, by adjusting the Fermi level of graphene material from 0eV to 1eV, the cross-polarized reflection amplitude has a modulation depth of more than 85% in the working frequency band of 0.6THz-1.0THz, the co-polarized reflection amplitude is less than 0.21, and the amplitude and phase adjustment do not interfere with each other and have relative independence. The design of graphene metasurface unit has the advantages of wide bandwidth, high modulation depth and amplitude and phase independence, which makes the design goals such as far-field multi-beam separation and arbitrary number change from algorithm optimization to practical application. Attached Figure Description

[0020] Figure 1This is a flowchart of the design method of the present invention;

[0021] Figure 2 This is a schematic diagram of the far-field diffraction matrix of the present invention;

[0022] Figure 3 This is a diagram showing the combined amplitude and phase optimization results of the present invention;

[0023] Figure 4 This is a diagram illustrating the iteration error of the present invention;

[0024] Figure 5 This is a structural diagram of the metasurface unit designed for this invention;

[0025] Figure 6 This is a graph showing the variation of the amplitude of the metasurface unit of the present invention with the Fermi level;

[0026] Figure 7 This is a phase change diagram of the metasurface unit of the present invention;

[0027] Figure 8 This is a diagram showing the result of adjusting the 16-beam to 12-beam configuration of the present invention.

[0028] Figure 9 This is a schematic diagram of a 9-beam dynamically adjustable far-field target according to the present invention;

[0029] Figure 10 The diagram shows the results of dynamically adjusting the 9-beam configuration to 8, 7, and 6 beams according to the present invention. Detailed Implementation

[0030] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0031] like Figure 1 As shown, this embodiment of the invention provides a terahertz graphene tunable metasurface design method based on machine learning, including the following steps:

[0032] (1) Based on the intensity and position of the far-field diffraction beam, set the target matrix A and the target matrix B. Specifically, the far field of the target is composed of digital matrices, representing the number of metasurface arrays and the number of far-field diffraction pixels; then the main beam position matrix value is set to 1, and the other side lobe position matrix values ​​are set to 0.

[0033] (2) Construct the SA-GS model. Based on the target matrix A, use the SA-GS-based optimization model to obtain the optimal phase distribution. This includes the following steps:

[0034] (21) Randomly set the metasurface phase and calculate its corresponding far-field distribution. Set the initial temperature, cooling rate and other parameters for the simulated annealing algorithm.

[0035] (22) In each simulated annealing iteration, the far-field intensity under the current phase distribution is calculated using Fourier transform and compared with the set target matrix A. The phase is updated through random perturbation and its value is adjusted according to the Fourier transform. After each iteration, the temperature is reduced according to the cooling rate to reduce the exploration amplitude and accelerate convergence.

[0036] (23) The algorithm updates and displays the current far-field intensity and phase distribution, helps to observe the optimization process, outputs the optimal phase distribution, and verifies the optimization effect through holographic reconstruction.

[0037] (3) Establish the JAP-PSO model. Based on the target matrix B, use the JAP-PSO amplitude-phase joint optimization model to obtain the final corrected perturbation phase distribution and quantization amplitude distribution. This includes the following steps:

[0038] (31) Initialize the value of all elements of the metasurface amplitude distribution to 1, which means that the amplitude of all units in the initial state of the metasurface array has not changed, and the phase distribution is the result obtained by SA-GS model optimization.

[0039] (32) Using the particle swarm optimization algorithm, a JAP-PSO amplitude-phase joint optimization model is established, and the number of particles, the maximum number of iterations, and the phase perturbation term are set.

[0040] (33) The optimization objective of the JAP-PSO model is to minimize the mean square error (MSE) between the optimized far-field intensity and the target far-field intensity. This optimization process involves adjusting the amplitude and phase simultaneously. By setting the search space of the particles, the solution of each particle is optimized step by step to ensure that the far-field scattering matrix is ​​consistent with the target matrix B.

[0041] (33) After the model completes the amplitude and phase optimization, the quantization order is defined. By optimizing the amplitude index and the corresponding threshold, the amplitude can be effectively quantized, minimizing the impact of quantization on far-field scattering.

[0042] (34) Based on the quantized amplitude distribution and phase, recalculate the quantized optimized far-field intensity. By comparing the results before and after optimization, visualize the optimal amplitude set, phase distribution, quantization threshold, and final far-field optimization result. Evaluate the optimization effect by calculating the mean square error (MSE) between the quantized far-field and the target far-field to ensure the effectiveness of the optimization method.

[0043] This invention also provides a terahertz multi-beam tunable metasurface, comprising a bottom metal reflector, a middle polyimide dielectric layer, and a top layer composed of metal and graphene. The bottom metal reflector is made of gold, which has low loss in the terahertz band. The middle dielectric layer can be made of flexible polyimide material with a dielectric constant of 3.0-3.6 and a loss tangent of 0.0001-0.05. The top layer is an open-ring pattern composed of metal and graphene. Left-handed or right-handed circularly polarized waves are incident, and the geometric phase of the open-ring pattern is obtained by rotating it by any angle from 0° to 180°, achieving 360° phase coverage. The metasurface unit period is 90μm-150μm, the gold thickness is 0.1μm-0.2μm, the dielectric thickness is 20μm-60μm, the outer radius of the open ring is 40μm-60μm, the ring width is 3μm-12μm, and the opening angle is 2°-12°.

[0044] Within the 0.6THz-1THz range, using circularly polarized wave incident light, the cross-polarized reflection amplitude is higher than 0.74, while the co-polarized reflection amplitude is lower than 0.10. By rotating the patch angle between 0° and 180°, a phase difference of twice the rotation angle is achieved. By changing the Fermi level of the top graphene material from 0eV to 1eV, the cross-polarized reflection amplitude is reduced to 0.10, a decrease of 86.40%, while maintaining the co-polarized reflection amplitude below 0.20.

[0045] The optimized metasurface units are arranged according to the optimal phase distribution. The Fermi level of the graphene material at the top of the unit structure is set according to the optimal quantization amplitude distribution. By using circularly polarized wave incident, the number, position and intensity of far-field beams can be arbitrarily changed.

[0046] Example 1: This embodiment of the invention provides a 16-beam dynamically tunable metasurface and its design method, including the following steps:

[0047] (1) The design process of a 16-beam adjustable metasurface is as follows: Figure 1 As shown, firstly, a set of target far-field matrices A and B are set, and then a phase optimization model based on SA-GS is constructed. After obtaining the optimal phase distribution, a joint amplitude and phase optimization model based on JAP-PSO is constructed to calculate the quantization amplitude distribution and perturbation phase distribution. Finally, the designed metasurface unit structure is used to perform array arrangement and far-field simulation verification.

[0048] (2) Using a circularly polarized wave as incident, set a set of target far-field matrices, such as Figure 2As shown, both target matrices A and B are 40×40, indicating that the far-field array consists of 40×40 metasurface units. At the center of matrix A, a 4×4 array of 16 beams is set, with the main beam intensity set to 1 and the remaining diffracted beam intensities set to 0. Similarly, at the center of matrix B, the four 1s in the upper right corner are changed to 0, thus reducing the number of beams to 12 through the amplitude-phase optimization model.

[0049] (3) Construct a phase calculation and optimization model based on SA-GS machine learning.

[0050] For the target far-field matrix A, the initial phase φ0(x, y) of the metasurface is randomly set and combined with the set target far-field matrix F(x, y) to form the far-field wave function f(x, y). A two-dimensional Fourier transform is performed on f(x, y) to obtain the wave function g(u, v) on the output metasurface plane, converting the time-domain spectrum to the frequency-domain spectrum. The phase of g(u, v) is combined with the amplitude distribution G(u, v) of the metasurface to form a new wave function g'(u, v), and an inverse Fourier transform is performed on it to convert it to the time-domain spectrum, obtaining the wave function f'(x, y) for the next iteration. The calculation formula is: replace φ0(x, y) with the phase of f'(x, y) and repeat the above process. The iterated phase φ1(x, y) is used as the initial phase for optimization in the SA algorithm. During each cooling process, a new temperature and phase distribution can be generated, calculated using the following formula:

[0051]

[0052] Where i is the number of iterations of the SA algorithm, T is the temperature, and Rand(M×N) is a randomly generated matrix between 0 and 1. In the calculation, the error function E is defined. i The difference between the target intensity and the theoretical far-field intensity of the output is expressed as:

[0053]

[0054] The smaller the error, the better the phase calculation result. The Metropolis criterion is used to accept the error, and the acceptance probability is defined as:

[0055]

[0056] As can be seen, with the increase of the number of iterations, the temperature gradually decreases, the probability of accepting errors increases, and the results tend to stabilize. Finally, the optimized phase distribution based on the target matrix A is obtained, as follows: Figure 3 As shown in (a).

[0057] (4) Construct an amplitude-phase joint optimization model based on JAP-PSO machine learning

[0058] Having obtained the optimal phase distribution for the far-field target matrix A using SA-GS, the next step is to construct a joint amplitude and phase optimization model for the far-field target matrix B. This model aims to minimize the mean square error (MSE) between the theoretical far-field intensity distribution and the target far-field B by controlling the amplitude and phase of each pixel, i.e., each metasurface unit in the array. To achieve this goal, the optimization model is constructed as follows:

[0059] First, the complex amplitude of the hologram is defined as:

[0060]

[0061] In the formula and Let x and y represent the amplitude and phase after iterative optimization, respectively, and let x and y represent the row and column of the pixel, respectively. By performing a two-dimensional Fourier transform (FFT2(⋅)) on this complex amplitude, the far-field intensity distribution can be obtained:

[0062]

[0063] The goal of model optimization is to minimize the error between the calculated far-field intensity distribution and the target far-field distribution. Let the target far-field be I. target The objective function is then optimized as follows:

[0064]

[0065] In the first optimization phase, we simultaneously optimize the amplitude and phase using a particle swarm optimization algorithm. The optimization process for amplitude and phase is viewed as a high-dimensional search problem, where the particle's position represents a combination of amplitude and phase, and each particle in the swarm is represented by a vector containing both amplitude and phase information. The particle's position and velocity are updated using the following formulas:

[0066]

[0067]

[0068] Among them, a i (t) is the position of particle i in the t-th iteration, v i (t) is the velocity of the particle, p i Let r1 be the individual optimal position of the particle, g be the global optimal position, r1 and r2 be random numbers, c1 and c2 be acceleration constants, and ω be the inertia weight. After each particle position update, we calculate the theoretical far-field intensity I corresponding to the particle position. theory The mean squared error (MSE) at that position is calculated. Since the particle's goal is to minimize the MSE, the fitness function for each particle is the MSE value corresponding to its current position. When the iteration error variation is less than 10... -6When the time is reached, optimization stops, and the perturbation phase distribution can be obtained as follows: Figure 3 As shown in (b), and the continuous amplitude distribution as shown in Figure 3 As shown in (c).

[0069] Following continuous amplitude optimization, the second optimization stage, quantization optimization, is performed. Quantization aims to convert the continuously optimized amplitude and phase into a finite number of discrete values, thus corresponding to the reflection amplitude of the actual metasurface unit. The quantization process uses a particle swarm optimization algorithm to calculate the index and threshold of the quantized amplitude. Assume the set of quantized amplitudes is A. Q If the quantization level is L, then the optimization objective is to select L quantization amplitudes such that the optimized far-field intensity is as close as possible to the target far-field intensity. The selection of quantization amplitudes depends on the amplitude of the optimized result and the quantization index S. idx Furthermore, the quantization threshold T is used to control the segmentation of each quantization level:

[0070]

[0071] Among them, A Q (S idx ) is the corresponding index S in the amplitude set. idx The quantization amplitude.

[0072] The goal of quantization optimization is also to minimize the mean square error. The error function after quantization is:

[0073]

[0074] I quant This is represented as the quantized far-field intensity distribution.

[0075] After the second stage of optimization is completed, the quantized amplitude distribution is obtained as follows: Figure 3 As shown in (d), the mean square error of the MSE of the JAP-PSO model is as follows: Figure 4 As shown in (a), after 405 iterations, the error in the last round is 2.5 × 10⁻⁶. -5 ,like Figure 4 As shown in (b).

[0076] (5) Design as follows Figure 5 The terahertz graphene metasurface unit shown is composed of three layers. Figure 5 In the diagram, 1 represents the metal backplate, which ensures total reflection of terahertz waves; 2 represents the dielectric layer used to transmit electromagnetic waves. 3 represents the metal open-ended ring, and 4 represents graphene material; together, they form the open-ended ring. The metal used has an electrical conductivity of 4.561 × 10⁻⁶. 7The gold dielectric layer is made of flexible polyimide material with a dielectric constant of 3.1 and a loss tangent of 0.05. The graphene material is initialized at 293K, has a relaxation time of 0.1ps, a thickness of 1nm, and a Fermi level of 0eV-1eV.

[0077] After optimization of the scanning parameters, the cell period p is 100 μm, the flexible polyimide layer thickness h1 is 48 μm, h0 is 0.2 μm, the inner radius r of the open ring is 35 μm, the outer radius R is 40 μm, the central angle θ corresponding to the gold portion is 225°, and the opening angle α of the open ring is 2°. Initially, the open ring is symmetrical along a 45° line. By rotating the open ring around its center, a phase difference can be constructed, thus covering the entire period.

[0078] The amplitude and phase characteristics of the graphene metasurface unit structure were calculated using circularly polarized wave incident radiation. Figure 6 (a) and (b) show the characteristics of cross-polarization amplitude variation in the range of 0.0THz-2.0THz; Figure 6 (c) and (d) show that within the 0.6THz-1.0THz operating frequency band with a 50% relative bandwidth, by changing the Fermi level of graphene, the modulation depth of cross-polarization is greater than 80%, and the co-polarization reflection amplitude is less than 0.28; the relationship between unit phase and Fermi level is as follows: Figure 7 As shown in (a), it can be seen from the figure that the maximum phase change does not exceed 20° when the Fermi level is changed. Figure 7 (b) shows a rotating open-ring pattern, with the phase completely covering 360°. Therefore, the designed terahertz graphene metasurface unit has dynamic tunable characteristics, and the amplitude and phase are independent and do not interfere with each other. This structural design meets the requirements of far-field dynamic control.

[0079] (6) After optimization, the quantized amplitude and phase distributions are obtained. The amplitudes are quantized to 0.10, 0.18, 0.23, 0.28, 0.33, 0.39, 0.45, 0.53, 0.62, and 0.74. The 0eV-1eV Fermi level values ​​corresponding to the amplitudes are then introduced into the graphene material. The phases are quantized to eight orders: 0°, 45°, 90°, 135°, 180°, 225°, 270°, and 315°. The open-ring pattern of each unit structure is rotated by the corresponding angle to form a far-field array, such as... Figure 8 As shown in (a).

[0080] (7) After the array arrangement is completed, a circularly polarized wave is used for incident radiation. When all Fermi levels of the graphene material are 0 eV, the far-field diffraction exhibits 16 beams, such as... Figure 8 As shown in (b); when the graphene Fermi level is set according to the model optimization results, the far-field diffraction dynamically changes to 12 beams, as... Figure 8As shown in (c); we can dynamically switch clockwise sequentially, and the far-field diffraction is as follows: Figure 8 As shown in (d), (e), and (f), the beam strength changed significantly, with a decrease of 8 dB, while the maximum difference in the main beam strength did not exceed 2.0 dB.

[0081] The main factor is the rise of the Fermi level, which causes graphene to gradually exhibit properties similar to metals, enhances its ability to manipulate charges, and results in a decrease in the reflection amplitude of the unit structure, which in turn leads to a decrease in the number of far-field beams.

[0082] Example 2: A graphene metasurface with arbitrary dynamic switching of 9 beams in the far field is provided, including the following steps.

[0083] (1) Define the far-field target matrix. The original far-field is as follows: Figure 9 As shown in (a), the 9 beams are evenly distributed; the dynamic control targets are set to 8, 7, and 6 beams respectively, as follows: Figure 9 As shown in (b), (c), and (d).

[0084] (2) The parameter settings of the algorithm model are the same as in Example 1. After JAP-PSO optimization, the final phase distribution and quantization amplitude distribution are obtained.

[0085] (3) By arranging the graphene metasurface units in an array according to the algorithm optimization results, and using circularly polarized waves as incident light, far-field simulation was performed, and the results are shown in Figure (10). Figure 10 (a) shows the original 9-beam setup without any adjustments. The dynamic adjustment results for beams 8, 7, and 6 are as follows: Figure 10 As shown in (b), (c), and (d), and Figure 9 The algorithm settings are basically the same. Therefore, the method proposed in this invention can achieve arbitrary switching of the number of far-field beams, and can be further extended to fields such as space communication and radar detection.

Claims

1. A machine learning-based method for designing tunable metasurfaces of terahertz graphene, characterized in that, Includes the following steps: (1) Based on the target intensity and position of the far-field diffracted beam, set the first target matrix and the second target matrix for phase optimization; wherein, the second target matrix is ​​used for amplitude and phase joint optimization; (2) Construct a phase optimization model based on the combination of simulated annealing and Fourier transform, and use the first objective matrix to obtain the optimal phase distribution of the metasurface through iterative optimization; (3) Construct a joint amplitude and phase optimization model based on particle swarm optimization, and use the second objective matrix to jointly optimize the amplitude and phase of the metasurface to obtain the corrected perturbation phase distribution and quantized amplitude distribution, thereby realizing the dynamic switching of the number of far-field beams.

2. The terahertz graphene tunable metasurface design method based on machine learning according to claim 1, characterized in that, In step (2), the phase optimization model uses the simulated annealing algorithm as a framework, combines Fourier transform to perform far-field calculation and phase update, and gradually approximates the optimal phase distribution corresponding to the target far field through random perturbation, temperature decay and error acceptance criteria.

3. The terahertz graphene tunable metasurface design method based on machine learning according to claim 1, characterized in that, In step (3), the amplitude and phase joint optimization model takes the particle swarm optimization algorithm as the core, takes the amplitude and phase as joint optimization variables, minimizes the error between the theoretical far field and the target far field through global search, and performs multi-level quantization on the amplitude after optimization to generate a discrete amplitude set that matches the control capability of the metasurface unit.

4. The terahertz graphene tunable metasurface design method based on machine learning according to claim 1, characterized in that, The amplitude and phase distributions obtained by the method can be adapted to a variety of tunable materials. By customizing the quantization order and threshold, it can meet the requirements of different hardware platforms and control precision.

5. A terahertz multi-beam tunable metasurface, characterized in that, The method described in any one of claims 1-4 is used to achieve the metasurface, which is composed of multiple periodically arranged graphene metasurface units; each unit includes a bottom metal reflector, an intermediate flexible dielectric layer, and a top open ring pattern composed of metal and graphene; phase modulation is achieved by rotating the angle of the open ring pattern, and amplitude modulation is achieved by adjusting the Fermi level of the graphene material, with amplitude and phase modulation being independent of each other.

6. The terahertz multi-beam tunable metasurface according to claim 5, characterized in that, By changing the Fermi level, the reflection amplitude can be continuously adjusted over a wide range while maintaining a stable phase response; by rotating the open ring pattern, phase coverage from 0° to 360° can be achieved, meeting the requirements of multi-beam dynamic control.

7. A terahertz multi-beam tunable metasurface according to claim 5, characterized in that, The metasurfaces are arranged in an array according to the optimal phase distribution and quantized amplitude distribution output by the optimization model. By setting the Fermi level and the rotation angle of the open ring of each unit graphene, the number, position and intensity of far-field diffraction beams can be dynamically switched arbitrarily.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the method described in any one of claims 1-4.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-4.

Images