Method for analyzing and designing multi-frequency wave-absorbing characteristics of terahertz electromagnetic metamaterial

Abstract

A terahertz electromagnetic metamaterial multi-frequency wave-absorbing characteristic analysis design method based on deep learning and structural parameters comprises the steps that (1) a structural prototype and a parameterized model of a terahertz electromagnetic metamaterial are established, and an electromagnetic metamaterial data set containing terahertz multi-frequency wave-absorbing characteristic information is constructed through iterative simulation calculation; 2) segmenting the data set into a 80% training data set and a 20% test data set; 3) establishing bidirectional spectrum networks SN-1 and SN-2, sending the training data set into the SN-1 and SN-2 for training, and verifying the performance through the test data set; 4) inputting the training data set into SN-1 and SN-2 network training, and testing the training effect by using the test data set; 5) utilizing the multivariate Gaussian distribution curve to randomly generate double-peak, three-peak and four-peak spectrum curve data as a multi-peak wave absorbing curve needing to be designed; and 6) inputting spectral curve data into an SN-2 network for retrieval, reconstructing a terahertz electromagnetic metamaterial multi-frequency wave-absorbing structure in CST and performing simulation calculation, and designing an electromagnetic metamaterial structure with a multi-frequency wave-absorbing characteristic.

Description

Technical Field

[0001] The present invention belongs to the field of analysis and design of multi-band terahertz electromagnetic metamaterial absorbers, and specifically relates to an analysis and design method of multi-band terahertz electromagnetic metamaterial absorbers based on deep learning and structural parameters. Background Art

[0002] At present, terahertz electromagnetic metamaterial absorbers are an important functional device that can uniformly absorb incident electromagnetic waves and play an important role in infrared imaging, filters, electromagnetic stealth and other fields. The emergence of multi-band terahertz electromagnetic metamaterial absorbers solves the problem of narrow absorption band of previous metamaterial absorbers, but its analysis and design process still has the problem of low efficiency. At present, deep learning has achieved great success in the field of information technology. In recent studies, deep learning methods have also been used in the analysis and design process of electromagnetic metamaterials, which greatly improves the analysis and design efficiency of electromagnetic metamaterial functional devices, making this method have important application value in the analysis, design and application of electromagnetic metamaterial functional devices.

[0003] Chinese patent CN115482893A relates to an electromagnetic metamaterial design method based on deep learning and structural variables, which inputs the electromagnetic metamaterial forward prediction data set into the forward prediction network model, and inputs the electromagnetic metamaterial reverse design data set into the reverse design network model, and then performs iterative cycle training for the forward prediction and reverse design of the electromagnetic metamaterial functional characteristics. However, the working frequency of the data used in this patent is limited to the microwave frequency band, lacks research results in the terahertz frequency band, and the patent lacks research content on the analysis and design of the multi-frequency characteristics of electromagnetic metamaterials.

[0004] The document "Gao F, Zhang Z, Xu Y, et al. Deep-learning-assisted designing chiralterahertz metamaterials with asymmetric transmission properties[J]. Journal of the Optical Society of America B-Optical Physics, 2022, 39(6): 1511-1519" discloses a deep neural network model that combines a spectral network (SN) and an extended network (EN), which realizes the analysis and design of asymmetric transmission properties of chiral electromagnetic metamaterials in the terahertz frequency band. However, the work in this document focuses on the analysis and design of asymmetric transmission properties using deep learning methods, lacks research on wave absorption properties, and lacks analysis and design work on the multi-frequency properties of electromagnetic metamaterials. Summary of the invention

[0005] Technical issues to be solved

[0006] In order to solve the problem of lack of analysis and design methods for the multi-frequency wave-absorbing characteristics of terahertz electromagnetic metamaterials and overcome the shortcomings of the prior art, the present invention provides a deep learning method combined with numerical simulation technology for the analysis and design process of the multi-frequency wave-absorbing characteristics of terahertz electromagnetic metamaterials.

[0007] Technical Solution

[0008] A method for analyzing and designing the multi-frequency wave-absorbing characteristics of terahertz electromagnetic metamaterials based on deep learning. The core of the method lies in the construction of a data set and the establishment of a neural network model. According to the structural characteristics of electromagnetic metamaterials with multi-frequency wave-absorbing characteristics, a data set containing multi-frequency wave-absorbing information is constructed. Then, two bidirectional spectrum neural networks SN-1 and SN-2 are constructed for the relevant performance parameters of the electromagnetic metamaterial wave-absorbing characteristics, which are used to realize the forward prediction and reverse design processes respectively. Finally, the trained neural network model is used to realize the on-demand reverse design process of the multi-frequency wave-absorbing characteristics of terahertz electromagnetic metamaterials.

[0009] A method for analyzing and designing multi-frequency wave absorption characteristics of terahertz electromagnetic metamaterials based on deep learning, comprising the following steps:

[0010] Step 1: Establish a structural prototype and parameterized model of terahertz electromagnetic metamaterials, and construct an electromagnetic metamaterial data set containing terahertz multi-frequency wave absorption characteristic information through iterative simulation calculations. Preferably, the steps are as follows:

[0011] Step 1.1: Construct a structural prototype and a parameterized model of the terahertz electromagnetic metamaterial, wherein the parameterized model includes the size parameter information of the electromagnetic metamaterial.

[0012] Step 1.2: Use MATLAB and electromagnetic simulation software CST electromagnetic simulation software to call the parameterized model in step 1.1 for simulation and calculate the electromagnetic response corresponding to the model. For the wave absorption characteristics of terahertz electromagnetic metamaterials, the relevant electromagnetic response can be expressed by reflectivity R, transmittance T and wave absorption rate A. In CST, these three parameters can be expressed by S parameters, that is, R = |S 11 | 2 、T=|S 21 | 2 、A=1-RT=1-|S 11 | 2 -|S 21 | 2 Here, it is stipulated that the bottom of the electromagnetic metamaterial absorber is a complete metal backplane, so that the transmission coefficient S 21 = 0, the relationship between the absorbance and the S parameter can be simplified to the reflection coefficient S 11 So the electromagnetic response here refers to the reflection coefficient S 11 and the absorption rate A.

[0013] Step 1.3: Combine the size parameter information in step 1.1 and the reflection coefficient S in step 1.2 11 The absorbance A is stored in a text file, which is a sample.

[0014] Step 1.4: Change the size parameter information in the parameterized model in step 1.1 and perform simulation calculations again. Repeat this process continuously to obtain all electromagnetic metamaterial samples and construct the entire data set. The samples in the data set include the size parameters, operating frequency range, and reflection coefficient S of the electromagnetic metamaterial. 11 and the wave absorption rate A. (It should be noted that in the process of changing the parameterized model size information, the structure of the electromagnetic metamaterial will gradually change from a single complete resonant structure to a combination of multiple resonant structures. In this way, the data set contains information on multi-frequency wave absorption characteristics so that the neural network can learn).

[0015] Step 2: Split the dataset constructed in step 1 into 80% training dataset and 20% test dataset.

[0016] Step 3: For the reflection coefficient S mentioned in step 1 11and absorbance A, two bidirectional spectrum neural networks SN-1 and SN-2 were built respectively, and the training data set was sent to SN-1 and SN-2 for training. After the training was completed, the performance of the two neural network models was verified by the test data set. Finally, it was verified that SN-1 established the electromagnetic metamaterial size parameters and reflection coefficient S 11 SN-2 establishes a mapping relationship between the electromagnetic metamaterial size parameter and the wave absorption rate A. Preferably, the steps are as follows:

[0017] Step 3.1: In the forward path of SN-1 and SN-2, the 6×1 size parameter will be converted into a 100×1 response spectrum, which means that the dimension of the input data is much lower than the dimension of the output data, making the convergence of the neural network full of challenges, especially when the spectrum curve changes drastically near the resonant frequency, the performance of the neural network is even more unpredictable. To solve this problem, a two-stage strategy is used to construct an approximate input-bottleneck-output network structure. Specifically: the input data is first up-dimensioned through the neural tensor layer, and then down-dimensioned through the fully connected neural network module to form a bottleneck layer. This module is called the downsampling module. After the training of this module is completed, it is frozen. The data flow enters the tensor dimension reshaping of the next module, the transposed convolution layer and the fully connected layer, and the dimension of the data flow gradually increases until the entire regression process is completed. This module is called the upsampling module. After the upsampling module is trained, the two pre-trained modules of downsampling and upsampling are combined and trained as the overall forward path of SN-1 and SN-2, and the forward prediction neural network is fine-tuned. This network structure increases the depth and training weights of the forward path, making it easy for the neural network to converge. Among them, the output of the neural tensor layer can be described as:

[0018] Output vector = activation function (transpose of input structure parameter vector × weight tensor W × input structure parameter vector + weight matrix V × input structure parameter vector + bias vector B)

[0019] In the formula, the activation function is a variant of the rectified linear unit (Relu) activation function, Leaky Relu; the dimension of the input structure vector is 6×1; the dimension of the output vector is set to 50×1; the dimension of the weight tensor W is 50×6×6, the dimension of the weight matrix V is 50×6, and the dimension of the bias vector B is 50×1. Step 3.2: In the reverse path of SN-1 and SN-2, three convolutional neural networks and one fully connected neural network module are included. Since there are multiple sets of different structural parameters in the reverse design process that may cause the same response spectrum, it is difficult for a single reverse neural network to converge. Therefore, according to the series neural network training strategy in previous work, the trained forward network in SN-1 and SN-2 is used as an auxiliary network to participate in the training of the reverse network. The overall loss function of the reverse network is set as the weighted average of the forward network loss and the separate reverse network loss:

[0020] The overall loss of the reverse network = mean square error loss function [forward neural network (reverse neural network (input spectrum vector))] + penalty coefficient × mean square error loss [reverse neural network (input spectrum vector)]

[0021] The mean square error loss function represents the average of the squares of the differences between the true value and the predicted value of the neural network. Step 4: Input the training data set segmented in step 2 into the SN-1 and SN-2 neural networks constructed in step 3 for training; use the test data set segmented in step 2 to test the training effect of the SN-1 and SN-2 neural networks. After the training is completed, two samples are randomly selected from the test data set to test the effects of the forward path and reverse path of the SN-1 and SN-2 neural networks.

[0022] Step 5: Using the multivariate Gaussian distribution curve, randomly generate double-peak, triple-peak, and quadruple-peak spectral curve data, and use them as the multi-peak absorption curve to be designed.

[0023] Step 6: Input the randomly generated double-peak, triple-peak, and quadruple-peak spectrum curve data in step 5 into the reverse path of the SN-2 neural network trained in step 3 for retrieval, and then obtain the size parameter data of the structure. According to the retrieved size parameters, reconstruct the corresponding terahertz electromagnetic metamaterial multi-frequency absorbing structure in CST and perform simulation calculations to obtain the corresponding absorption rate curves of these structures. According to the results, it can be proved that the data set and neural network model constructed based on the above steps can reversely design the electromagnetic metamaterial structure with multi-frequency absorbing characteristics according to the number of absorption peaks required.

[0024] The present invention also provides a terahertz electromagnetic metamaterial absorber with multi-frequency wave absorbing characteristics, which is designed and manufactured using the above-mentioned analysis and design method of the multi-frequency wave absorbing characteristics of terahertz electromagnetic metamaterials based on deep learning and structural parameters.

[0025] Beneficial Effects

[0026] According to the physical mechanism of multi-frequency absorbing electromagnetic metamaterials, the present invention proposes a parametric modeling method for generating electromagnetic metamaterial samples containing multi-frequency absorbing structural information, and then constructing the entire data set. Next, the present invention establishes a deep learning model for the analysis and design process of multi-frequency absorbing characteristics of terahertz electromagnetic metamaterials. The model includes two bidirectional deep neural networks SN-1 and SN-2. After training, SN-1 successfully establishes 6 size parameters and reflection coefficient S 11 SN-2 successfully established the mapping relationship between the six size parameters and the absorption rate A. The loss functions of the forward path and reverse path of the two neural networks converged to 10 -3 and 10 -2 In addition, in order to further evaluate the prediction and design capabilities of the deep neural network model, two samples were randomly selected from the test data set. On the one hand, SN-1 and SN-2 were used for forward prediction. The results showed that for the reflection coefficient S 11 , the predicted curve of SN-1 is in good agreement with the real curve calculated by CST. For the absorption rate A, the predicted curve of SN-2 is in good agreement with the real curve calculated by CST, which shows that SN-1 and SN-2 have good prediction capabilities. On the other hand, using SN-1 and SN-2 for reverse design, the results show that the dimensional parameters designed by SN-1 and SN-2 are very close to the real dimensional parameters, and the shape of the reconstructed electromagnetic metamaterial structure diagram is also very consistent, which shows that SN-1 and SN-2 have good reverse design capabilities.

[0027] The present invention utilizes a multivariate Gaussian distribution curve to randomly generate bimodal, trimodal, and quadrumodal spectral curve data, and uses it as the multimodal absorption curve to be designed. Then, the randomly generated bimodal, trimodal, and quadrumodal spectral curve data are input into the reverse path of the SN-2 neural network for retrieval, thereby obtaining the size parameter data of the structure, and based on the retrieved size parameters, the corresponding terahertz electromagnetic metamaterial multi-frequency absorption structure is reconstructed in CST and simulated, and the absorption rate curves corresponding to these structures are obtained. The results show that the electromagnetic metamaterial absorber modeled and simulated according to the size parameters designed by the neural network can be designed with dual-frequency, triple-frequency, and quadruple-frequency absorption characteristics in the working frequency band according to the requirements of the number of design peaks. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 A schematic diagram of the structural prototype of the terahertz electromagnetic metamaterial absorber is shown.

[0029] Figure 2The schematic diagram of the prototype of electromagnetic metamaterial metal structure and the schematic diagram of the metal structure of the dataset sample are shown:

[0030] (a) Schematic diagram of the prototype of electromagnetic metamaterial metal structure;

[0031] (b)-(e) Schematic diagrams of metal structures of dataset samples.

[0032] Figure 3 The flowchart of electromagnetic metamaterial parameterized modeling calculation and data set sample generation is shown.

[0033] Figure 4 Shown is the overall structure diagram of the deep neural network model.

[0034] (a) Internal structure diagram of the bidirectional deep neural network models SN-1 and SN-2;

[0035] (b) Schematic diagram of the composition of the Neural Tensor layer.

[0036] Figure 5 The loss function graphs of deep neural network models SN-1 and SN-2 are shown:

[0037] (a) Loss function of the upsampling module in the SN-1 forward path;

[0038] (b) SN-1 reverse path loss function;

[0039] (c) Loss function of the upsampling module in the forward path of SN-2;

[0040] (d) SN-2 reverse path loss function.

[0041] Figure 6 The forward path prediction results of deep neural network models SN-1 and SN-2 are shown in the figure:

[0042] (a) SN-1 predicts the reflection coefficient S11 of random sample 1;

[0043] (b) SN-1 predicts the reflection coefficient S11 of random sample 2;

[0044] (c) SN-2 predicts the absorption rate A of random sample 1;

[0045] (d) SN-2 predicts the absorption rate A of random sample 2.

[0046] Figure 7 The reverse path design results of deep neural network models SN-1 and SN-2 are shown:

[0047] (a) Comparison of design size and actual size of random sample 1;

[0048] (b) Comparison diagram of the designed structure and the actual structure of random sample 1;

[0049] (c) Comparison of design size and actual size of random sample 2;

[0050] (d) Comparison of the designed structure and the actual structure of random sample 2.

[0051] Figure 8 The figure shows the on-demand design process and result diagram of the deep neural network model SN-2 for realizing the multi-frequency wave-absorbing characteristics of terahertz electromagnetic metamaterials. DETAILED DESCRIPTION

[0052] The present invention will be described in detail below with reference to the accompanying drawings and specific implementation modes, but the present invention is not limited thereto.

[0053] A method for designing multi-frequency wave-absorbing characteristics of terahertz electromagnetic metamaterials based on deep learning, comprising the following steps:

[0054] Step 1: Establish the structural prototype of the terahertz electromagnetic metamaterial absorber and the corresponding parameterized model;

[0055] Step 2: Use MATLAB and electromagnetic simulation software CST to continuously call the established parameterized model and perform numerical simulation calculations to generate samples of terahertz electromagnetic metamaterial absorbers until the data set is completed;

[0056] Step 3: Split the constructed dataset into 80% training dataset and 20% test dataset;

[0057] Step 4: A two-stage strategy is used to construct the forward path of the SN-1 and SN-2 neural networks. The forward path includes a neural tensor layer network, a fully connected neural network, and a transposed convolutional network. The entire network structure is an approximate input-bottleneck-output network structure.

[0058] Step 5: Construct the reverse path of the SN-1 and SN-2 neural networks. The reverse path includes three convolutional neural networks and one fully connected neural network. According to the training strategy of the tandem neural network, the trained forward path network is used as an auxiliary network to participate in the training of the reverse path network. The overall loss of the reverse path network is set as the weighted average of the loss of the forward network and the loss of the separate reverse network.

[0059] Step 6: Input the training data set segmented in step 3 into the SN-1 and SN-2 neural networks constructed in steps 4-5 for training; use the test data set segmented in step 3 to test the training effect of the SN-1 and SN-2 neural networks.

[0060] Step 7: Using the multivariate Gaussian distribution curve, randomly generate double-peak, triple-peak, and quadruple-peak spectral curve data, and use them as the multi-peak absorption curve to be designed.

[0061] Step 8: Input the double-peak, triple-peak, and quadruple-peak spectral curve data randomly generated in step 7 into the reverse path of the SN-2 neural network trained in step 6 for retrieval, and then obtain the size parameter data of the structure. According to the retrieved size parameters, reconstruct the corresponding terahertz electromagnetic metamaterial multi-frequency absorbing structure in CST, and perform simulation calculations to verify whether the absorption rate of these electromagnetic metamaterial structures meets the design requirements.

[0062] In a specific embodiment, the present invention provides a method for designing multi-frequency wave-absorbing properties of terahertz electromagnetic metamaterials based on deep learning:

[0063] (1) Figure 1 The structural prototype of the terahertz electromagnetic metamaterial absorber is shown. The periodic unit size of the structure is p = 30 μm. The materials of the top metal structure and the bottom metal plate are gold, and the conductivity is σ = 4.6×10 7 s / m, initial thickness t = 0.2μm; the size parameters of the top metal structure are l1 = 15μm, l2 = 9μm, l3 = 6μm, b = 3μm; the material of the middle dielectric layer substrate is polyimide polymer (Polyimide), with a relative dielectric constant of 3.5, a loss tangent value of 0.0027, and an initial thickness d = 1.8μm. The frequency domain solver of the commercial electromagnetic simulation software CST Microwave Studio is used to simulate and calculate the wave absorption characteristics of terahertz electromagnetic metamaterials in the frequency range of 0.1THz to 10.0THz.

[0064] Through simulation, we can get the reflection coefficient S of the terahertz electromagnetic metamaterial for the incident electromagnetic wave. 11 and the transmission coefficient S 21 , the wave absorption rate of electromagnetic metamaterial can be expressed by the reflection coefficient S 11 and the transmission coefficient S 21 Calculation: A = 1-|S 11 | 2 -|S 21 | 2 ; When designing the prototype of the electromagnetic metamaterial absorber, a metal backplane was designed at the bottom layer to prevent the transmission of electromagnetic waves, making the transmission coefficient S 21 = 0, so the absorption rate can be determined only by the reflection coefficient S 11 Calculation: A = 1-|S 11 | 2 .

[0065] (2) According to the structural prototype of the terahertz electromagnetic metamaterial absorber, the design size parameters are selected, including the length of the metal rod l1, l2, l3, the width of the metal rod b, the thickness of the metal layer t, and the thickness of the dielectric layer d. By designing and screening these sizes, all electromagnetic metamaterial unit structure samples (a total of 36,750 cases) are generated. Table 1 shows the size sampling intervals in the parameterized space of the electromagnetic metamaterial unit structure. Among all the collected samples, 80% of the samples are assigned as training data sets, and the remaining 20% ​​of the samples are assigned as test data sets.

[0066] Table 1. Design parameters of electromagnetic metamaterial unit structure

[0067]

[0068]

[0069] Figure 2 The metal prototype of the electromagnetic metamaterial absorber structure and the sample structure in the data set are shown. It can be seen that the metal prototype of the electromagnetic metamaterial absorber structure is a complete structure with four-fold rotational symmetry. Figure 2 (a). When the design size parameters are constantly changing, the data set can collect not only samples with a single complete structure, but also samples with multiple resonant structures rotated, fused or arranged in the same plane, see Figure 2 (b)-(e). In this way, as the design size parameters continue to change within the specified parameter space, the types and number of samples in the data set continue to increase, and eventually a sample data set containing information on a single resonant structure and multiple resonant structures can be obtained to achieve the analysis and design of the multi-frequency wave absorption characteristics of electromagnetic metamaterials.

[0070] Figure 3 The flowchart of the parametric modeling simulation and data set construction process is shown. First, the parametric model program file is called through MATLAB to sample the variation space of the specific size parameters of the electromagnetic metamaterial; then, MATLAB is used to call the electromagnetic simulation software CST Microwave Studio, and the terahertz electromagnetic metamaterial absorber is modeled in the CST software, and then numerical simulation is performed to calculate the reflection coefficient S within the working frequency range. 11 , and then according to the reflection coefficient S 11 The relationship between the absorption rate A and the absorption rate is used to calculate the absorption rate; finally, the size parameters, frequency range, and reflection coefficient S 11 , the absorption rate A is stored in a text file, which is a sample. After continuous iterations, the number and types of samples in the data set continue to increase, and finally a sample data set containing information on a single resonant structure and multiple resonant structures is obtained, laying the foundation for the analysis and prediction of multi-frequency absorption characteristics and on-demand design using deep neural networks.

[0071] (3) Establish a deep neural network model based on deep learning. Figure 4 The overall framework of the deep neural network model is presented, which includes the reflection coefficient S 11 and the absorption rate A to design two bidirectional spectrum neural networks SN-1 and SN-2, see Figure 4 (a) Evaluation of electromagnetic metamaterial structure and reflection coefficient S using the forward and reverse paths of SN-1 11 The non-intuitive relationship between the electromagnetic metamaterial structure and the wave absorption rate A is evaluated by using the forward path and reverse path of SN-2. A two-stage strategy is used to construct the forward path of the SN-1 and SN-2 neural networks. The forward path includes a neural tensor layer network, a fully connected neural network, and a transposed convolutional network. The entire network structure is an approximate input-bottleneck-output network structure; among them, the neural tensor layer is Figure 4 (b), its output can be described as:

[0072] Output vector = activation function (transpose of input structure parameter vector × weight tensor W × input structure parameter vector + weight matrix V × input structure parameter vector + bias vector B)

[0073] In the formula, the activation function is a variant of the rectified linear unit (Relu) activation function, Leaky Relu; the dimension of the input structure vector is 6×1; the dimension of the output vector is set to 50×1; the dimension of the weight tensor W is 50×6×6, the dimension of the weight matrix V is 50×6, and the dimension of the bias vector B is 50×1.

[0074] The reverse path of the SN-1 and SN-2 neural networks includes three convolutional neural networks and one fully connected neural network. According to the training strategy of the tandem neural network, the trained forward path network is used as an auxiliary network to participate in the training of the reverse path network. The overall loss of the reverse path network is set as the weighted average of the loss of the forward network and the loss of the separate reverse network:

[0075] The overall loss of the reverse network = mean square error loss function [forward neural network (reverse neural network (input spectrum vector))] + penalty coefficient × mean square error loss [reverse neural network (input spectrum vector)]

[0076] The mean square error loss function represents the average value of the square of the difference between the true value and the predicted value of the neural network.

[0077] (4) Input the segmented training data set into the constructed SN-1 and SN-2 neural networks for training; use the segmented test data set to test the training effect of the SN-1 and SN-2 neural networks. Figure 5(a)-(b) shows the change curves of the forward path and reverse path loss functions of SN-1, and Figure 5(c)-(d) shows the change curves of the forward path and reverse path loss functions of SN-2. The loss functions of the forward path and reverse path of these two neural networks converge to 10 -3 and 10 -2 After training, two samples are randomly selected from the test data set to test the effects of the forward and reverse paths of the SN-1 and SN-2 neural networks. Figure 6 (a)-(b) shows the reflection coefficient S of these two samples using SN-1. 11 The predicted results are compared with the actual results. Figure 6 (c)-(d) shows the comparison between the predicted results and the actual results of the wave absorption rate A of the two samples using SN-2. The results show that the predicted curve and the actual curve are in good agreement, indicating that SN-1 and SN-2 have good prediction capabilities. Figure 7 (a) shows the comparison between the size parameters obtained by SN-1 and SN-2 for the reverse design of random sample 1 and the actual size parameters, and Figure 7 (b) shows the electromagnetic metamaterial structure diagram reconstructed according to the design parameters and the actual parameters; Figure 7 (c) shows the comparison between the size parameters obtained by SN-1 and SN-2 for the reverse design of random sample 1 and the actual size parameters, and Figure 7 (d) shows the electromagnetic metamaterial structure diagram reconstructed according to the design parameters and the actual parameters. The results show that the values ​​of the size parameters designed by SN-1 and SN-2 and the actual size parameters are very close, and the shapes of the reconstructed electromagnetic metamaterial structure diagrams are also very consistent, indicating that SN-1 and SN-2 have good reverse design capabilities.

[0078] (5) Using the multivariate Gaussian distribution curve, randomly generate double-peak, triple-peak, and quadruple-peak spectral curve data, and use them as the multi-peak absorption curve to be designed. Figure 8 (a) shows the three curves.

[0079] (6) The randomly generated double-peak, triple-peak, and quadruple-peak spectrum curve data are input into the reverse path of the SN-2 neural network for retrieval, and then the size parameter data of the structure is obtained. Based on the retrieved size parameters, the corresponding terahertz electromagnetic metamaterial multi-frequency absorbing structure is reconstructed in CST and simulated and calculated to obtain the corresponding absorption rate curves of these structures. The whole process is as follows Figure 8 As shown, Figure 8 (a) shows the bimodal, trimodal, and quadrumodal spectral curves randomly generated using multivariate Gaussian distribution curves; Figure 8 (b) shows the Figure 8 (a) The dimensional parameters obtained by inputting the spectral curve shown into SN-2 for reverse design; Figure 8 (c) shows the electromagnetic metamaterial structure reconstructed according to the design size and the wave absorption rate curve calculated by CST simulation. It can be found that the electromagnetic metamaterial absorber modeled and simulated according to the size parameters designed by the neural network can design terahertz electromagnetic metamaterial absorbers with dual-frequency, triple-frequency, and quad-frequency wave absorption characteristics in the working frequency band according to the requirements of the number of design peaks.

[0080] In summary, according to the physical mechanism of the multi-frequency absorber, a parametric modeling method is proposed to generate electromagnetic metamaterial samples containing multi-frequency absorbing structural information. Then, based on a large number of training samples, the inventors proposed a deep learning-based method for the analysis and design of multi-frequency absorbing characteristics. The deep neural network model includes two bidirectional networks SN-1 and SN-2, which are used to simulate the reflection coefficient S of the terahertz electromagnetic metamaterial. 11 and the absorbance A. After training, the deep neural network model can accurately establish the non-intuitive mapping relationship between the electromagnetic metamaterial structural parameters and the full electromagnetic spectrum. Finally, through the multivariate Gaussian distribution curve, multi-peak spectral curve data is randomly generated as the absorbance curve of the desired design. These curves are input into the deep learning model for reverse design, and the designed electromagnetic metamaterial absorber shows corresponding multiple absorption peaks.

[0081] It should be noted that the embodiments described above are only preferred embodiments of the present invention. For those skilled in the art, without departing from the principles of the present invention, the present invention may be modified, improved and replaced with equivalents, and these modifications, improvements and equivalent replacements are also considered to fall within the scope of protection of the claims of the present invention.

Claims

1. An analysis and design method for multi-frequency wave absorption characteristics of terahertz electromagnetic metamaterials based on deep learning and structural parameters, comprising: Step 1: Establish a structural prototype and parameterized model of terahertz electromagnetic metamaterials, and construct an electromagnetic metamaterial data set containing terahertz multi-frequency wave absorption characteristics through iterative simulation calculations; Step 2: Split the dataset constructed in step 1 into 80% training dataset and 20% test dataset; Step 3: Build two bidirectional spectral neural networks SN-1 and SN-2 respectively, send the training data set to SN-1 and SN-2 for training. After the training is completed, verify the performance of the two neural network models through the test data set; Step 4: Input the training data set segmented in step 2 into the SN-1 and SN-2 neural networks built in step 3 for training; use the test data set segmented in step 2 to test the training effect of the SN-1 and SN-2 neural networks. After the training is completed, randomly select two samples from the test data set to test the SN-1 and SN-2 neural networks; Step 5: Using the multivariate Gaussian distribution curve, randomly generate double-peak, triple-peak, and quadruple-peak spectrum curve data, and use them as the multi-peak absorption curve to be designed; Step 6: Input the double-peak, triple-peak, and quadruple-peak spectral curve data randomly generated in step 5 into the SN-2 neural network trained in step 3 for retrieval, and then obtain the size parameter data of the structure. According to the retrieved size parameters, reconstruct the corresponding terahertz electromagnetic metamaterial multi-frequency absorbing structure in CST and perform simulation calculations to obtain the corresponding absorption rate curves of these structures, so as to design electromagnetic metamaterial structures with multi-frequency absorbing characteristics as needed.

2. The method for analyzing and designing multi-frequency wave absorption characteristics of terahertz electromagnetic metamaterials based on deep learning and structural parameters according to claim 1 is characterized in that: In step 1, the parameterized model includes dimensional parameter information of the electromagnetic metamaterial.

3. The method for analyzing and designing multi-frequency wave absorption characteristics of terahertz electromagnetic metamaterials based on deep learning and structural parameters according to claim 1 or 2, characterized in that: In step 1, the parameterized model is called for simulation through MATLAB and electromagnetic simulation software CST electromagnetic simulation software to calculate the electromagnetic response corresponding to the parameterized model; the parameterized model and the electromagnetic response are used to perform iterative simulation calculations to construct a data set.

4. The method for analyzing and designing multi-frequency wave absorption characteristics of terahertz electromagnetic metamaterials based on deep learning and structural parameters according to claim 3 is characterized in that: The electromagnetic response refers to the reflection coefficient S 11 and the absorption rate A.

5. According to any one of claims 1 to 4, the analysis and design method of multi-frequency wave absorption characteristics of terahertz electromagnetic metamaterials based on deep learning and structural parameters, the samples in the data set include the size parameters, working frequency range, reflection coefficient S of the electromagnetic metamaterial 11 and the absorption rate A.

6. The method for analyzing and designing multi-frequency wave absorption characteristics of terahertz electromagnetic metamaterials based on deep learning and structural parameters according to claim 1, characterized in that: In step 3, a two-stage strategy is used to construct the forward path of the SN-1 and SN-2 neural networks. The forward path includes a neural tensor layer network, a fully connected neural network, and a transposed convolutional network. The entire network structure is an approximate input-bottleneck-output network structure; the reverse path of the SN-1 and SN-2 neural networks is constructed. The reverse path includes three convolutional neural networks and one fully connected neural network. According to the training strategy of the series neural network, the trained forward path network is used as an auxiliary network to participate in the training of the reverse path network. The overall loss of the reverse path network is set as the weighted average of the loss of the forward network and the loss of the separate reverse network.

7. The method for analyzing and designing multi-frequency wave absorption characteristics of terahertz electromagnetic metamaterials based on deep learning and structural parameters according to claim 6 is characterized in that: The input data is first upgraded through the neural tensor layer, and then reduced through the fully connected neural network module to form a bottleneck layer, called the downsampling module. After the downsampling module is trained, it is frozen, and the data flow enters the next module for tensor dimension reshaping, transposing the convolution layer and the fully connected layer. The dimension of the data flow gradually increases until the entire regression process is completed, which is called the upsampling module. After the upsampling module is trained, the downsampling module and the upsampling module are combined and trained as the overall forward path of SN-1 and SN-2 to fine-tune the forward prediction neural network.

8. The method for analyzing and designing multi-frequency wave absorption characteristics of terahertz electromagnetic metamaterials based on deep learning and structural parameters according to claim 6 or 7, characterized in that: The output of the neural tensor layer is described as: Output vector = activation function (transpose of input structure parameter vector × weight tensor W × input structure parameter vector + weight matrix V × input structure parameter vector + bias vector B) In the formula, the activation function is a variant of the rectified linear unit (Relu) activation function, Leaky Relu; the dimension of the input structure vector is 6×1; the dimension of the output vector is set to 50×1; the dimension of the weight tensor W is 50×6×6, the dimension of the weight matrix V is 50×6, and the dimension of the bias vector B is 50×1.

9. The method for analyzing and designing multi-frequency wave absorption characteristics of terahertz electromagnetic metamaterials based on deep learning and structural parameters according to claim 6, characterized in that: In step 3, the overall loss of the reverse path network is set to be the weighted average of the loss of the forward network and the loss of the individual reverse networks: The overall loss of the reverse network = mean square error loss function [forward neural network (reverse neural network (input spectrum vector))] + penalty coefficient × mean square error loss [reverse neural network (input spectrum vector)] The mean square error loss function represents the average value of the square of the difference between the true value and the predicted value of the neural network.

10. A terahertz electromagnetic metamaterial absorber with multi-frequency wave absorption characteristics, characterized in that: The terahertz electromagnetic metamaterial is designed and manufactured using the analysis and design method for multi-frequency wave absorption characteristics of terahertz electromagnetic metamaterials based on deep learning and structural parameters as described in any one of claims 1 to 9.

Images