Terahertz metasurface difunctional adjustable polarization converter based on vanadium dioxide

Abstract

The invention discloses a vanadium dioxide-based terahertz metasurface dual-function adjustable polarization converter, which comprises a converter body, and is characterized in that the converter body is composed of sub-wavelength periodic units arranged periodically; wherein the sub-wavelength periodic unit comprises a pattern layer, a dielectric layer and a reflecting layer; the bottom end of the pattern layer is fixedly connected with the top end of the dielectric layer; the pattern layer is of a similar I-shaped piece structure and comprises a first patch, a second patch and a third patch; the two opposite ends of the first patch are fixedly connected with the two ends of the second patch and the two ends of the third vanadium dioxide patch respectively. The design speed and precision of the dual-function polarization converter are effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of terahertz polarization converters, and in particular to a terahertz metasurface dual-function adjustable polarization converter based on vanadium dioxide. Background Art

[0002] A terahertz polarization converter is a device used to change the polarization state of electromagnetic waves in the terahertz band. Polarization refers to the characteristic of the vibration direction of an electromagnetic wave, and the role of a polarization converter is to change the amplitude of the components of the incident wave in two mutually perpendicular directions so that they meet the requirements of linear polarization and circular polarization, respectively. Polarization converters can improve the performance and flexibility of terahertz systems. In imaging, communication, and sensing technologies, they can be used to enhance signal quality and improve the resolution and sensitivity of the system. In addition, terahertz polarization converters are typically designed based on artificial microstructures or new materials to achieve broadband and efficient polarization conversion. With the development of materials science and micro-nanotechnology, terahertz polarization converters will continue to play an increasingly important role in scientific research and industry.

[0003] Convolutional neural networks are good at processing high-dimensional data. Big data usually involves a large amount of unstructured information, which is difficult to be effectively processed by traditional analysis methods. CNN automatically extracts hierarchical features from data through pooling layers and convolution layers, reducing the computational burden, reducing dependence on manual feature engineering, and improving the efficiency and accuracy of analysis.

[0004] However, conventional terahertz metasurface dual-function tunable polarization converters have the following disadvantages:

[0005] (1) The metal resonant structure used in conventional terahertz metasurface dual-function tunable polarization converters causes high loss and dispersion problems;

[0006] (2) Traditional terahertz polarization converter design relies on experience or single-objective optimization methods, which has low practicality. Summary of the Invention

[0007] Purpose of the invention: The purpose of the present invention is to provide a terahertz metasurface dual-function adjustable polarization converter based on vanadium dioxide to solve the problems existing in the background technology.

[0008] Technical solution: The vanadium dioxide-based terahertz metasurface dual-function adjustable polarization converter described in the present invention includes a converter body, which is composed of periodically arranged subwavelength periodic units; wherein the subwavelength periodic units include a pattern layer, a dielectric layer and a reflective layer; the bottom end of the pattern layer is fixedly connected to the top end of the dielectric layer, and the bottom end of the dielectric layer is fixedly connected to the top end of the reflective layer; the pattern layer is an I-shaped structure, including: a first patch, a second patch and a third patch; the opposite ends of the first patch are respectively fixedly connected to the second patch and the two ends of the vanadium dioxide third patch.

[0009] Furthermore, the I-shaped structure is symmetrically arranged about two diagonals of the sub-wavelength periodic unit.

[0010] Furthermore, the period p of the sub-wavelength periodic unit is 100-110 μm; the side length l of the pattern layer is 60-80 μm; the width d1 of both sides of the first patch is 2-3 μm; the interval d2 on the outer diagonal line of the first patch is 4-6 μm; the interval d3 on the inner diagonal line of the first patch is 12-20 μm; and the middle width d5 ​​of the first patch (211) is 10-20 μm.

[0011] Furthermore, the width d4 of the second patch is 5-10 μm; the width d6 of the third patch is 15-25 μm.

[0012] Furthermore, the thickness h1 of the dielectric layer is 35 μm; the thickness h2 of the pattern layer and the reflective layer are both 0.2 μm; the dielectric layer is made of polyimide, and the reflective layer is made of gold.

[0013] Furthermore, the material of the first patch is gold; the materials of the second patch and the third patch are vanadium dioxide. By adjusting the ambient temperature to change the conductivity of vanadium dioxide, dual-function switching of broadband linear polarization conversion and broadband circular polarization conversion is achieved.

[0014] The present invention provides a vanadium dioxide-based terahertz polarization converter optimization method for optimizing the polarization converter, comprising the following steps:

[0015] (1) Set the model parameters and set the optimization range and step size for the adjustable parameters l, d1, d2, d3, d4, and d5;

[0016] (2) Build a bidirectional convolutional neural network (CNN) model and set the hyperparameter grid;

[0017] (3) Input the simulation results into the CNN model for training and establish a bidirectional mapping between model parameters (X) and reflection coefficients (Y);

[0018] (4) Compare the model prediction value and the simulation actual value under the same design parameters;

[0019] (5) The constraint condition is that the highest comprehensive ranking of polarization conversion rate PCR and axial ratio AR is achieved;

[0020] (6) Output the optimization results and compare them with the actual simulation values ​​for verification.

[0021] Furthermore, in step (2), the bidirectional CNN model includes: forward propagation path (X→Y): extracting features through three layers of 1D convolution layers, with the number of convolution kernels being 64, 128, and 256 respectively, and each layer being followed by batch normalization (BatchNorm) and a maximum pooling layer (MaxPooling1D) with a step size of 2, and finally outputting a 700-dimensional reflection coefficient through a fully connected layer; backward propagation path (Y→X): reconstructing the 6-dimensional model parameters through a symmetrical deconvolution and upsampling structure;

[0022] The calculation formula of the convolutional neural network CNN model is as follows:

[0023]

[0024] O is the output matrix, I is the input matrix, w is the convolution kernel, k h , k w are the height and width of the convolution kernel, representing the size of the convolution kernel; the calculation formula of the LeNet-5 model is as follows:

[0025]

[0026] Among them, the size of the input matrix I is (H, W), and the size of the filter convolution kernel is (k h , k w ), the output matrix O has size (OH, OW), padding is P, and stride is S.

[0027] An electronic device according to the present invention includes a memory and a processor, wherein the memory stores a computer program, and the processor implements the steps of any one of the methods when executing the program.

[0028] The computer-readable storage medium of the present invention stores a computer program, which implements the steps of any one of the methods when executed by a processor.

[0029] Beneficial effects: Compared with the existing technology, the present invention has the following significant advantages: 1. The polarization converter enables the device to switch between broadband linear polarization and broadband circular polarization by changing the ambient temperature, thereby achieving the purpose of dual-function control;

[0030] 2. The present invention simplifies the structural optimization process by performing convolutional neural network (CNN) processing on the simulation data of structural parameters, thereby effectively improving the design speed and accuracy of the dual-function polarization converter. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the structure of the converter body of the present invention;

[0032] Figure 2 Schematic diagram of the architecture of the sub-wavelength periodic unit of the present invention;

[0033] Figure 3 is a side view of a sub-wavelength periodic unit of the present invention;

[0034] Figure 4 A top view of a sub-wavelength periodic unit of the present invention;

[0035] Figure 5 is a flow chart of the present invention;

[0036] Figure 6 The reflection coefficient diagram and PCR diagram of the vanadium dioxide of the present invention when it is in an insulating state;

[0037] Figure 7 PCR graphs of vanadium dioxide of the present invention at different conductivities;

[0038] Figure 8 This is a reflection coefficient diagram of the vanadium dioxide of the present invention when it is in a metallic state;

[0039] Figure 9 The reflection coefficient phase diagram and phase difference diagram of the vanadium dioxide of the present invention when it is in a metallic state;

[0040] Figure 10 This is an axial ratio diagram of the vanadium dioxide of the present invention when it is in a metallic state;

[0041] Figure 11 A comparison chart of the simulation value and the predicted value of the present invention;

[0042] In the figure: 1, converter body; 2, sub-wavelength periodic unit; 21, pattern layer; 22, dielectric layer; 23, reflective layer; 211, first patch; 212, second patch; 213, third patch. DETAILED DESCRIPTION

[0043] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0044] like Figures 1-4As shown, the embodiment of the present invention provides a terahertz metasurface dual-function adjustable polarization converter based on vanadium dioxide, including a converter body 1 and a bidirectional neural network optimization method. The converter body 1 is composed of a plurality of sub-wavelength periodic units 2 arranged periodically. The sub-wavelength periodic units 2 each include a pattern layer 21, a dielectric layer 22 and a reflective layer 23. The bottom end of the pattern layer 21 is fixedly connected to the top end of the dielectric layer 22, and the bottom end of the dielectric layer 22 is fixedly connected to the top end of the reflective layer 23. The pattern layer 21 is an I-shaped part. The pattern layer 21 includes a first patch 211, a group of second patches 212 and a group of third patches 213. The opposite ends of the first patch 211 are respectively connected to the second patch (212) and the third patch 213. The two ends of the three patches 213 are fixedly connected; the shape of the graphic composed of the first patch 211, the second vanadium dioxide patch 212 and the third vanadium dioxide patch 213 is similar to the "I" structure, and is symmetrically arranged about the two diagonals of the square periodic unit. The middle dielectric layer 22 is made of polyimide, which ensures that the phase relationship of the terahertz wave during propagation meets the requirements of polarization conversion by cumulatively adjusting the phase of the terahertz wave. The bottom reflection layer 23 is made of gold, which reflects the terahertz wave that passes through the dielectric layer 22. Through reflection, a multiple interference cavity structure is formed inside the system. This result greatly improves the polarization conversion efficiency of the entire system through continuous wave reflection and phase control. The material of the first patch 211 is gold; the material of the second patch 212 and the third patch 213 is vanadium dioxide. By adjusting the ambient temperature to change the conductivity of vanadium dioxide, dual-function switching of broadband linear polarization conversion and broadband circular polarization conversion is achieved.

[0045] The I-shaped structure formed by bonding several first patches 211, several second patches 212, and a third patch 213 is the pattern layer 21. The period p of the subwavelength periodic unit 2 is 100-110 μm, the side length l of the pattern layer 21 is 60-80 μm, the width d1 of the first patch 21 is 2-3 μm, the spacing d2 on the outer diagonal of the first patch 211 is 4-6 μm, the spacing d3 on the inner diagonal of the first patch 211 is 12-20 μm, and the center width d5 ​​of the first patch 211 is 10-20 μm. The patch width d4 of the second patch 212 is 5-10 μm, and the width d6 of the third patch 213 is 15-25 μm. The thickness h1 of the dielectric layer 22 is 35 μm, and the thickness h2 of the pattern layer 21 and the reflective layer 23 is 0.2 μm.

[0046] like Figure 5 As shown, the embodiment of the present invention also provides a method for optimizing a terahertz polarization converter based on vanadium dioxide, comprising the following steps:

[0047] Step 1. Set converter parameters: Set the model and parameters, and set the optimization range and step size for the adjustable parameters l, d1, d2, d3, d4, and d5 of the polarization converter;

[0048] Step 2: Set model parameters: Set the hyperparameter grid of the convolutional neural network (CNN) model; set the calculation formula of the convolutional neural network (CNN) model as follows:

[0049]

[0050] O is the output matrix, I is the input matrix, w is the convolution kernel, k h , k w are the height and width of the convolution kernel, representing the size of the convolution kernel; the calculation formula of the LeNet-5 model is as follows:

[0051]

[0052] Among them, the size of the input matrix I is (H, W), and the size of the filter convolution kernel is (k h , k w ), the output matrix O has size (OH, OW), padding is P, and stride is S.

[0053] Step 3, simulation learning: input the simulation results into the CNN model for learning; simulation learning is based on the basic principles of the above-mentioned convolutional neural network, and is specially optimized for the inverse design problem of terahertz metasurface devices. On the basis of the aforementioned convolution operation formula, the preferred technical solution of the present invention adopts a bidirectional CNN architecture to realize the bidirectional mapping of the polarization converter model parameters (X) and the reflection coefficient (Y) obtained after CST simulation. Forward propagation path (X→Y): Features are extracted through three layers of 1D convolution, and the number of kernels is set to 64 / 128 / 256 respectively. Each layer of convolution is followed by BatchNorm and MaxPooling1D, and the step size is set to 2. Finally, a 700-dimensional reflection coefficient is output through the fully connected layer; Backward propagation path (Y→X): A symmetrical structure is used to reconstruct the 6-dimensional model parameters through deconvolution and upsampling. In particular, in the design of the convolution layer, the preferred technical solution of the present invention strictly follows the convolution operation rules: the output feature map size is obtained by the formula OH=(H+2P-k h ) / S+1 is calculated, where H is the input length, P is the padding, and k h is the kernel size, and S is the stride (in this model, S=1 and P=1). This design ensures that the temporal dimension of each layer's output remains stable during feature extraction, thereby providing regular input features for the subsequent fully connected layers. The model uses MSE as the loss function and is trained end-to-end using the Adam optimizer. This maintains the powerful feature extraction capabilities of traditional CNNs while achieving accurate bidirectional prediction of metasurface device parameters.

[0054] Step 4: Compare the numerical values: Compare the simulation results with the actual values ​​under the same design parameters.

[0055] Step 5: Set the constraint conditions: The constraint condition is that the comprehensive ranking of the polarization conversion rate PCR and the axial ratio AR is the highest;

[0056] Step 6: Output the results: Obtain the optimized results and compare them with the actual simulation values ​​to obtain the final optimization results.

[0057] In the present invention, a reflective temperature-controlled dual-function terahertz polarization conversion device is designed. The device realizes different functions under different temperature conditions. When the vanadium dioxide is in a metallic state, that is, at a temperature of 68 degrees Celsius or above, the device realizes line-to-line polarization conversion. When the vanadium dioxide is in an insulating state, the device realizes line-to-circular polarization conversion, thereby achieving the purpose of actively controlling the polarization conversion function. The electromagnetic response characteristics of the metamaterial microstructure used are simulated by the three-dimensional electromagnetic simulation software CST Microwave Studio, and the metamaterial reflection spectra under different conductivities are screened. The period p of the subwavelength periodic unit 2 is 100-110 μm, the side length l of the pattern layer 21 is 60-80 μm, the width d1 of the two sides of the first patch 21 is 2-3 μm, the spacing d2 on the outer diagonal line of the first patch 211 is 4-6 μm, the spacing d3 on the inner diagonal line of the first patch 211 is 12-20 μm, and the center width d5 ​​of the first patch 211 is 10-20 μm. The patch width d4 of the second patch 212 is 5 to 10 μm, and the width d6 of the third patch 213 is 15 to 25 μm. The thickness h1 of the dielectric layer 22 is 35 μm, and the thickness h2 of the pattern layer 21 and the reflective layer 23 is 0.2 μm. When an x-direction linearly polarized terahertz wave is incident perpendicularly along the -z axis onto the surface of the terahertz polarization converter, the conversion rate PCR of the polarization converter is expressed as:

[0058]

[0059] Among them, r yx represents the cross-polarization reflection coefficient of the incident x-ray polarized wave, r xx It represents the reflection coefficient of the co-polarized wave when the x-ray polarized wave is incident. When the vanadium dioxide is in a metallic state through temperature adjustment, circular polarization will occur. At this time, when the y-line polarized wave is incident, the axial ratio is:

[0060]

[0061] m=|r xx | 4 +|r yx | 4 +2|r xx | 2 |r yx | 2 cos(2△φ),

[0062] Wherein, Δφ is the phase difference between the co-polarization reflection coefficient and the cross-polarization reflection coefficient; m is an intermediate variable. In the present invention, the plasma frequency of the vanadium dioxide material satisfies

[0063]

[0064] Where σ = 1.4x10 15 By adjusting the conductivity from 200S / M to 2x10 5 The S / M temperature range is 23 to 87 degrees Celsius, which enables dynamic regulation of vanadium dioxide. The collision frequency satisfies

[0065]

[0066] where ε ∞ =12,γ=5.57×10 13 rad / s, cross-polarization and co-polarization curves for vanadium dioxide in an insulating state with a conductivity of 200 S / m. When the temperature causes the vanadium dioxide to be in an insulating state, the polarization conversion ratio is greater than 0.9 in the range of 0.666-1.566 THz. The circular polarization axial ratio curve for the insulating state of vanadium dioxide with a conductivity of 200,000 S / m satisfies the requirement for an axial ratio less than 3 dB in the range of 0.589-1.026 THz. By adjusting the vanadium dioxide conductivity, the terahertz polarization converter based on vanadium dioxide can switch between linear and circular polarization.

[0067] The present invention provides a vanadium dioxide-based terahertz polarization converter optimization method, comprising the following steps: setting a model and parameters, setting an optimization range and step size for adjustable parameters l, d1, d2, d3, d4, and d5 of the polarization converter; setting a hyperparameter grid of a convolutional neural network (CNN) model; inputting simulation results into the CNN model for learning; comparing the results obtained by model simulation under the same design parameters with the actual values; setting a constraint condition, wherein the constraint condition is that the comprehensive ranking of the polarization conversion rate (PCR) and the axial ratio (AR) is the highest; obtaining an optimized result, and comparing it with the actual simulation value to obtain a final optimization result. The optimization ranges of l, d1, d2, d3, d4, and d5 are shown in Table 1 below (unit: μm).

[0068] Table 1 Convolutional neural network (CNN) model input model parameter settings

[0069]

[0070] like Figure 6 The figure shows the linear polarization conversion performance of vanadium dioxide (VO2) in the insulating state of the present invention. Specifically, Figure 6Figure (a) shows the cross-polarization (ryx) and co-polarization (rxx) reflection coefficient curves of the present invention after reflection when the incident wave is an x-polarized wave. Figure (b) shows the corresponding polarization conversion rate (PCR) curve. As shown in Figure (a), within the operating frequency band of 0.686THz to 1.60THz, the amplitude of the cross-polarization reflection coefficient is greater than 0.8, the co-polarization reflection coefficient is less than 0.3, and at the three resonant frequencies of 0.732, 1.116 and 1.50THz, the co-polarization coefficient is approximately equal to 0. As shown in Figure (b), within the broadband range of 0.666THz to 1.566THz, the polarization conversion rate is greater than 0.9, the relative bandwidth is 80.6%, and the polarization conversion rate is approximately equal to 1 at the three resonant frequencies of 0.732, 1.116 and 1.50THz, proving that the present invention has high linear polarization conversion performance in the insulating state.

[0071] like Figure 7 The figure shows the performance of the polarization conversion rate (PCR) of the present invention at different conductivities. This figure shows that by changing the conductivity of vanadium dioxide from 200 S / m to 200,000 S / m, the polarization conversion performance of the device can be effectively controlled. When the conductivity of vanadium dioxide is 200 S / m, the polarization conversion rate is greater than 0.9 in the broadband range of 0.666 THz to 1.566 THz, and the relative bandwidth is 80.6%, which can achieve efficient broadband linear polarization conversion; when the conductivity is 3000 S / m, efficient linear polarization conversion in multiple bands can be achieved, with bandwidth ranges of 0.641-0.744 THz, 1.056-1.152 THz and 1.416-1.574 THz, respectively. Within the bandwidth range, the polarization conversion rate is greater than 0.9, and the relative bandwidths are 14.8%, 8% and 10.5%, respectively; when the conductivity is 200,000 S / m, the polarization conversion rate is greater than 0.9 in the range of 1.062 THz to 1.118 THz, and the relative bandwidth is 2.5%. It is explained that by regulating the conductivity of vanadium dioxide, the device can achieve dynamic switching and flexible control of linear polarization conversion functions from high-efficiency broadband, multi-band to specific narrowband.

[0072] like Figures 8 to 10 The figure shows the polarization conversion performance of vanadium dioxide (VO2) in the metallic state. Figure 8 The cross-polarization (ryx) and co-polarization (rxx) reflection coefficients of vanadium dioxide (VO2) in the metallic state are demonstrated. In the range of 0.616THz to 1.004THz, the absolute value of the amplitude difference between cross-polarization and co-polarization is within 0.3, indicating that within this range, the amplitudes of the cross-polarization and co-polarization reflection coefficients are close, which is one of the prerequisites for achieving circular polarization conversion.

[0073] Figure 9(a) shows the reflection phase of cross-polarization (ryx) and co-polarization (rxx), Figure 9 (b) shows the phase difference between the two And between 0.51THz and 0.962THz, the phase difference is stable at around -90° and 270°, which is a left-hand circularly polarized wave and meets the requirements of the phase difference of the orthogonal component.

[0074] Figure 10 The axial ratio (AR) curve calculated using the axial ratio formula is presented. The results show that within the frequency band of 0.589THz to 1.026THz, the axial ratio (AR) is less than 3dB, and the relative bandwidth is 54.11%, demonstrating that the device can serve as an efficient broadband circularly polarized converter when vanadium dioxide (VO2) is in the metallic state.

[0075] like Figure 11 As shown in the figure, it is a comparison diagram of the verification results of the optimization method proposed in the present invention. The figure compares the predicted value (Prediction) of the bidirectional convolutional neural network model with the simulation value (Simulation) after simulation by electromagnetic simulation software. The comparison value used is the polarization conversion rate (PCR) calculated by the polarization conversion rate formula of the cross-polarization and co-polarization obtained by the optimization model. Figure 11 (a) shows the comparative results for a special case, when the structural parameters to be predicted exceed the parameter boundaries during model training. The prediction performance obtained by the optimized model is shown in the figure. In this case, the predicted values ​​have large errors in both amplitude and bandwidth. Figure 11 (b) shows the prediction comparison within the parameter boundary. The prediction curve and the simulation curve show a high degree of consistency in the entire working frequency band, which fully proves that the bidirectional neural network model constructed by the present invention has extremely high prediction accuracy within the preset optimization parameter space and successfully verifies the effectiveness of the optimization method.

Claims

1. A terahertz metasurface dual-function adjustable polarization converter based on vanadium dioxide, comprising a converter body (1), characterized in that: The converter body (1) is composed of periodically arranged sub-wavelength periodic units (2); wherein the sub-wavelength periodic units (2) include a pattern layer (21), a dielectric layer (22) and a reflective layer (23); the bottom end of the pattern layer (21) is fixedly connected to the top end of the dielectric layer (22), and the bottom end of the dielectric layer (22) is fixedly connected to the top end of the reflective layer (23); the pattern layer (21) is an I-shaped structure, comprising: a first patch (211), a second patch (212) and a third patch (213); the opposite ends of the first patch (211) are fixedly connected to the two ends of the second patch (212) and the third vanadium dioxide patch (213), respectively.

2. The vanadium dioxide-based terahertz metasurface dual-function adjustable polarization converter according to claim 1, characterized in that: The I-shaped structure is symmetrically arranged about two diagonal lines of the sub-wavelength periodic unit (2).

3. The vanadium dioxide-based terahertz metasurface dual-function adjustable polarization converter according to claim 1, characterized in that: The period p of the sub-wavelength periodic unit (2) is 100-110 μm; the side length l of the pattern layer (21) is 60-80 μm; the width d1 of both sides of the first patch (211) is 2-3 μm; the interval d2 on the outer diagonal line of the first patch (211) is 4-6 μm; the interval d3 on the inner diagonal line of the first patch (211) is 12-20 μm; and the middle width d5 ​​of the first patch (211) is 10-20 μm.

4. The vanadium dioxide-based terahertz metasurface dual-function adjustable polarization converter according to claim 1, characterized in that: The width d4 of the second patch (212) is 5 to 10 μm; the width d6 of the third patch (213) is 15 to 25 μm.

5. The vanadium dioxide-based terahertz metasurface dual-function adjustable polarization converter according to claim 1, characterized in that: The thickness h1 of the dielectric layer (22) is 35 μm; the thickness h2 of the pattern layer (21) and the reflective layer (23) are both 0.2 μm; the dielectric layer (22) is made of polyimide, and the reflective layer (23) is made of gold.

6. The vanadium dioxide-based terahertz metasurface dual-function adjustable polarization converter according to claim 1, characterized in that: The material of the first patch (211) is gold; the materials of the second patch (212) and the third patch (213) are vanadium dioxide. By adjusting the ambient temperature to change the conductivity of the vanadium dioxide, dual-function switching of broadband linear polarization conversion and broadband circular polarization conversion is achieved.

7. A terahertz polarization converter optimization method based on vanadium dioxide, used to optimize the polarization converter, characterized in that: The following steps are involved: (1) Set the model parameters and set the optimization range and step size for the adjustable parameters l, d1, d2, d3, d4, and d5; (2) Build a bidirectional convolutional neural network (CNN) model and set the hyperparameter grid; (3) Input the simulation results into the CNN model for training and establish a bidirectional mapping between model parameters (X) and reflection coefficients (Y); (4) Compare the model prediction value and the simulation actual value under the same design parameters; (5) The constraint condition is that the highest comprehensive ranking of polarization conversion rate PCR and axial ratio AR is achieved; (6) Output the optimization results and compare and verify them with the actual simulation values.

8. The method for optimizing a terahertz polarization converter based on vanadium dioxide according to claim 7, characterized in that: In step (2), the bidirectional CNN model includes: forward propagation path (X→Y): extracting features through three layers of 1D convolution layers, with the number of convolution kernels being 64, 128, and 256 respectively, and each layer being followed by batch normalization (BatchNorm) and a maximum pooling layer (MaxPooling1D) with a step size of 2, and finally outputting a 700-dimensional reflection coefficient through a fully connected layer; backward propagation path (Y→X): reconstructing the 6-dimensional model parameters through a symmetrical deconvolution and upsampling structure; The calculation formula of the convolutional neural network CNN model is as follows: O is the output matrix, I is the input matrix, w is the convolution kernel, k h , k w are the height and width of the convolution kernel, representing the size of the convolution kernel; the calculation formula of the LeNet-5 model is as follows: Among them, the size of the input matrix I is (H, W), and the size of the filter convolution kernel is (k h , k w ), the output matrix O has size (OH, OW), padding is P, and stride is S.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the program, the steps of the method according to any one of claims 7 to 8 are implemented.

10. A computer-readable storage medium, characterized in that A computer program is stored, and when the program is executed by a processor, the steps of the method according to any one of claims 7 to 8 are implemented.

Images