A metasurface structure design method based on joint discriminative generative adversarial network

Abstract

The application relates to the field of electromagnetic super technology, in particular to a metasurface structure design method based on a joint discriminant generative adversarial network, a periodic arrangement of a metasurface unit structure and related parameters is designed, and is input into electromagnetic simulation software to obtain electromagnetic response, a forward prediction data set and a reverse design data set are constructed; the forward prediction data set and the reverse design data set are respectively divided into a training set and a test set; a forward prediction network model is constructed, the forward prediction data set is sent into a simulator network Simulator used for forward prediction to perform training, and the forward prediction network model parameters are fixed after the training is completed. The application adds a pre-trained forward prediction network and a discriminator network for discriminating spectral similarity into a reverse design network, so that the network can well understand the relationship between structure design and overall optical response, more attention is paid to the absorption spectrum of the generated structure, new design can be generated, and the application can be extended to analyze other metasurface structures.

Description

Technical Field

[0001] The present invention relates to the technical field of electromagnetic metamaterials, and in particular to a metasurface structure design method based on a joint discriminant generative adversarial network. Background Art

[0002] The concept of electromagnetic metamaterials, first proposed in 1999, refers to artificially manufactured, periodic, three-dimensional composite materials. Through the periodic arrangement of artificial structures, electromagnetic metamaterials can produce physical properties not possessed by naturally occurring materials, such as negative dielectric constant and negative magnetic permeability. An electromagnetic metasurface, an artificial layered material with a thickness less than the wavelength, is considered the two-dimensional counterpart of an electromagnetic metamaterial. Unlike electromagnetic metamaterials, electromagnetic metasurfaces utilize the sudden phase and amplitude changes caused by the sudden changes in electromagnetic waves on both sides of the metasurface to control the phase and amplitude distribution of the reflected and transmitted fields in space. Electromagnetic metasurfaces can flexibly control the amplitude, phase, and other characteristics of electromagnetic waves. Compared with traditional electromagnetic metamaterials, they have superior properties such as low loss and good ductility.

[0003] Super-unit surface design refers to the design of a surface structure for electromagnetic wave absorption within a specific frequency range, and is usually used in the preparation of absorbing materials or absorbing coatings. The absorption spectrum describes the absorption ability of a material or surface to electromagnetic waves of different frequencies. In super-unit surface design, there may be a variety of different surface structure designs, but the absorption spectra corresponding to these designs may tend to be similar or overlapping. Different surface designs may lead to similar absorption performance, which is a "many-to-one" relationship. The specific manifestation of this "many-to-one" relationship depends on factors such as the material, structure and preparation process used in the super-unit surface design. When designing a super-unit surface, these factors are usually adjusted to achieve the desired absorption performance, and it may be found that there is a certain degree of overlap or similarity between different designs.

[0004] Generative adversarial network (GAN) models are used for inverse design of supercell surfaces to provide cell structures with arbitrary patterns. GANs consist of a generative network that generates images and a discriminative network that distinguishes between generated images and real images. The generative network is trained to produce realistic images that deceive the discriminative network, and the discriminative network is trained not to be deceived by the generative network. The two networks compete with each other at each training step; ultimately, this competition leads to mutual improvement, allowing the generative network to produce realistic images of higher quality than when learning alone. To reduce time-consuming iterative simulations, data-driven design methods based on deep learning have been introduced in nanophotonics. Using generative adversarial networks to design nanophotonic structures is not restricted to predefined shapes. For a given input absorption spectrum, the network generates an ideal design in the form of an image. The conditional deep convolutional generative adversarial network (cDCGAN) algorithm, a recently developed algorithm to address the instability issues of GANs, relies not only on random noise vectors but also on additional conditional information. This conditional information can be of any type, such as category labels or text descriptions. By incorporating this conditional information, cDCGAN can control the samples generated by the generator to conform to the provided conditions.

[0005] When applying cDCGAN to metamaterial reverse design, conditional information can represent the design specifications or goals that the generated metamaterial structure must meet. Given a spectrum A and a latent vector z, the conditional information can be additional information describing the desired optical performance. The generator uses this conditional information to adjust the generated metamaterial structure, utilizing cDCGAN to generate metamaterial structures with specific optical properties, rather than simply generating structures from random noise to meet the required optical performance. The generated designs are presented as images, essentially providing any possible design that can meet the desired optical properties, without being restricted to specific structures. Overall, the cDCGAN model implements a framework for reverse designing metamaterial surface nanostructures. The network is trained through a game between a generator and a discriminator to meet the desired design goal. The generated results focus on the similarity between the generated image representing the metamaterial surface structure and existing samples in the training set, without effectively constraining the fit of the corresponding spectral response to the target response. Summary of the Invention

[0006] The purpose of the present invention is to overcome the above problems or at least partially solve the above problems, and proposes a super surface structure design method based on a joint discriminant generative adversarial network. By effectively introducing spectral response loss in the model architecture design and objective function design, the network back propagation process mainly relies on the degree of fit between the spectral response corresponding to the generated surface structure design and the target spectral response to optimize the network parameters, effectively avoiding the problem that traditional reverse design can only generate super unit surface structures in a similar image domain based on the spectrum.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for designing a hypersurface structure based on a joint discriminant generative adversarial network, comprising the following steps:

[0008] S1. Designing a periodic arrangement of electromagnetic metasurface unit structures and related parameters, and inputting them into electromagnetic simulation software to obtain electromagnetic responses, constructing forward prediction datasets and reverse design datasets; both the forward prediction datasets and the reverse design datasets include generated electromagnetic response absorption spectrum data and metasurface patterns;

[0009] S2. Split the forward prediction dataset and the reverse design dataset into training sets and test sets respectively, where the proportion of the training set is larger than the proportion of the test set;

[0010] S3. Build a forward prediction network model, send the segmented forward prediction data set to the simulator network Simulator for forward prediction for training, and fix the forward prediction network model parameters after training is completed;

[0011] S4. Construct a reverse design network model, including a generator network Generator for generating metasurface patterns, an image discriminator network Discriminator_A for determining the authenticity of patterns, and a spectral discriminator network Discriminator_B for determining the authenticity of spectra, and initialize the network parameters;

[0012] S5. Send the training set in the reverse-engineered dataset into the generator network Generator, send the generated image and the real image into the image discriminator network Discriminator_A, send the spectrum predicted by the generated image and the real spectrum into the spectrum discriminator network Discriminator_B, and perform iterative cycle training.

[0013] In a preferred embodiment, the electromagnetic metasurface unit includes an upper layer, a middle layer and a lower layer; the upper layer is divided into a coding pattern area; the middle layer dielectric substrate is made of F4B; and the lower layer is covered with full metal copper.

[0014] In a preferred embodiment, in S1, the steps of constructing a forward prediction dataset and a reverse design dataset include:

[0015] Draw different metasurface unit patterns, construct corresponding electromagnetic metasurface units, and input them into electromagnetic simulation software to obtain the corresponding electromagnetic response;

[0016] The encoding matrix of the metasurface unit pattern is used as the network input. In this process, the absorption rate Amp is combined with the reflectivity S 11 and transmittance S 12 The relationship is calculated as follows:

[0017]

[0018] By reflectivity S 11 and transmittance S 12 Calculate the absorption rate Amp, perform uniform frequency sampling on the absorption rate Amp, express it as a multi-dimensional vector, use it as the label of the data set, and construct a forward prediction data set;

[0019] The absorption rate Amp is represented as a multi-dimensional vector and used as input at the same time, and its corresponding metasurface unit matrix is ​​used as a label to construct a reverse design dataset.

[0020] In a preferred embodiment, in said S3, the forward prediction network model constructed, the simulator network Simulator adopts the ResNet18 structure, the optimizer of the network Simulator is set to Adam, and the learning rate is 1e-3;

[0021] use Represents the network loss function, where y is the true value, is the predicted value, m is the dimension, and the convergence of the model is improved through a dynamic adjustment mechanism. During the training process, the network's prediction of the absorption rate is gradually optimized. After the training is completed, the forward prediction network model parameters are fixed.

[0022] In a preferred embodiment, in the S4, a reverse design network model is constructed, wherein the generator network Generator is composed of a deconvolution layer, receives an absorbance vector as input, and generates a matrix representing the metasurface pattern through a deconvolution operation; the image discriminator network Discriminator_A is composed of a convolution layer, and the real image in the data set and the image generated by the generator and the absorption spectrum are input into the discriminator, and the image features are converted into digital features to output the discrimination probability of the authenticity of the image; the spectral discriminator Discriminator_B is composed of a linear layer, and inputs the real spectrum and the predicted spectrum of the generated image, and outputs the discrimination probability of the authenticity of the spectrum.

[0023] In a preferred embodiment, in S2, the forward prediction data set input is a 64×64 encoding matrix, and the corresponding label is a 51-dimensional absorption rate Amp vector; the reverse design data set input is a 51-dimensional absorption rate Amp vector, and the corresponding label is a 64×64 encoding matrix.

[0024] In a preferred embodiment, in S5, each round of iterative calculation process includes the following steps:

[0025] S5.1. Feed the absorption coefficient vector of the reverse-engineered dataset into the generator network. This input vector is concatenated with random noise and processed through multiple deconvolution layers and activation functions to generate a metamaterial structure image.

[0026] S5.2. The generated metamaterial structure image and the original metamaterial structure image corresponding to the absorption rate vector are fed into the image discriminator network Discriminator_A. The input vector is concatenated with random noise and processed through multiple convolutional layers and activation functions to output a probability for determining the authenticity of the image.

[0027] S5.3. The generated metamaterial structure image is fed into the forward prediction network. The predicted absorbance vector and the true absorbance vector are input into the spectrum discriminator Discriminator_B. After processing through multiple linear layers and activation functions, the output is used to judge the probability of the spectrum authenticity.

[0028] S5.4. Calculate the error through the image prediction value and the spectral prediction value, and perform the back propagation algorithm to update the network parameters of the real part generator network Generator, the image discriminator network Discriminator_A, and the spectral discriminator network Discriminator_B.

[0029] In a preferred embodiment, in S5.2, the loss of the image discriminator network Discriminator_A is defined as:

[0030]

[0031] Where m is the batch size, the superscript i represents the i-th data in the batch, the loss is the cross entropy between the image dataset x and the generated data G(A,z), and the discriminator output is between 0 and 1. In order to distinguish between real and fake pictures, D(x i ) approaches 1, D(G(A i ,z i )) approaches 0, and this loss defines the similarity of the pattern between the image data and the generated image.

[0032] In a preferred embodiment, in S5.3, the loss of the spectral discriminator Discriminator_B is defined as:

[0033]

[0034] The loss is measured by the spectral data A and the predicted spectrum The cross entropy of the discriminator is between 0 and 1. In order to distinguish true and false spectra, D(A i ) approaches 1, Approaching 0, this loss defines the similarity between the spectral data and the predicted spectrum.

[0035] In a preferred embodiment, in S5.4, the parameters of the generator are updated by backpropagation through the two discriminator losses:

[0036]

[0037] In the game between the generator and the discriminator, D(G(A i ,z i ))and It will approach 1, which introduces the reconstruction error of the spectral response. λ is a parameter that balances the spectrum and image loss, and λ = 0.1 is selected.

[0038] Compared to existing technologies, this invention offers the following advantages: By integrating a pretrained forward prediction network and a discriminator network for spectral similarity into the inverse design network, the network is able to better understand the relationship between the structural design and its overall optical response, placing greater emphasis on generating the structure's absorption spectrum. This network architecture is not limited to predefined structures but can also generate new designs and be extended to analyze other metasurface structures, making it widely applicable to the system design of nanophotonic structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 Schematic diagram of the structure of the electromagnetic metasurface unit of the present invention;

[0040] Figure 2 The structure and training flow chart of the forward network model of the present invention;

[0041] Figure 3 The structure and training flow chart of the reverse-engineered network model of the present invention;

[0042] Figure 4 Schematic diagram of the starting template of the electromagnetic metasurface unit of the present invention;

[0043] Figure 5 This is a graph showing the downward trend of the loss function during the forward network training process of the present invention;

[0044] Figure 6 To reverse design the network, find out the difference between the design value and simulation value of the existing pattern;

[0045] Figure 7 The difference between the design value and simulation value of the reverse design network for the new pattern. DETAILED DESCRIPTION

[0046] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0047] See also Figure 1-4 The present invention provides a technical solution: a method for designing a hypersurface structure based on a joint discriminant generative adversarial network, comprising the following steps:

[0048] S1. Designing a periodic arrangement of electromagnetic metasurface unit structures and related parameters, and inputting them into electromagnetic simulation software to obtain electromagnetic responses, constructing forward prediction datasets and reverse design datasets; both the forward prediction datasets and the reverse design datasets include generated electromagnetic response absorption spectrum data and metasurface patterns;

[0049] In specific implementation, the electromagnetic metasurface unit is as follows Figure 1 As shown, Figure 1 (a) is a front view, and (b) is a side view. The electromagnetic metasurface unit consists of an upper layer, a middle layer, and a lower layer. The upper layer is divided into a coding pattern area, which is divided into a 64×64 matrix. The coding pattern metal copper patch has a conductivity of 5.8e+007S / m and a thickness of t1 = 0.1mm. The middle layer dielectric substrate is made of F4B with a dielectric constant of 2.65(1+0.003i), a width L = 2mm, and a thickness t2 = 0.2mm. The lower layer is covered with all-metal copper, with a conductivity of 5.8e+007S / m and a thickness t3 = 0.1mm. The electromagnetic response of the electromagnetic metasurface unit in this structure depends on the coding pattern style of the upper layer.

[0050] In specific implementation, the steps for constructing the forward prediction dataset and the reverse design dataset are as follows:

[0051] Draw different 64×64 pixel metasurface unit patterns, construct corresponding electromagnetic metasurface units, and input them into electromagnetic simulation software to obtain the corresponding electromagnetic response. The frequency range is set to 120-160 Thz.

[0052] The 64×64 encoding matrix of the metasurface unit pattern is used as the network input. In this process, the absorption rate Amp is compared with the reflectivity S 11 and transmittance S 12 The relationship is calculated as follows:

[0053]

[0054] By reflectivity S 11 and transmittance S 12 Calculate the absorption rate Amp, perform uniform frequency sampling on the absorption rate Amp, express it as a multi-dimensional vector, use it as the label of the data set, and construct a forward prediction data set;

[0055] The absorption rate Amp is represented as a 51-dimensional vector and used as input, and its corresponding metasurface unit matrix is ​​used as a label to construct a reverse design dataset.

[0056] S2, split the forward prediction dataset and the reverse design dataset into 80% training set and 20% test set respectively;

[0057] In the specific implementation, the forward prediction data set input is a 64×64 encoding matrix, and the corresponding label is a 51-dimensional absorption rate Amp vector; the reverse design data set input is a 51-dimensional absorption rate Amp vector, and the corresponding label is a 64×64 encoding matrix.

[0058] S3. Build a forward prediction network model, send the segmented forward prediction data set to the simulator network Simulator for forward prediction for training, and fix the forward prediction network model parameters after training is completed;

[0059] In the specific implementation, the forward prediction network model constructed, the simulator network Simulator adopts the ResNet18 structure, the optimizer of the network Simulator is set to Adam, and the learning rate is 1e-3;

[0060] use Represents the network loss function, where y is the true value, is the predicted value, m is the dimension, and the convergence of the model is improved through a dynamic adjustment mechanism. During the training process, the network's prediction of the absorption rate is gradually optimized. After the training is completed, the forward prediction network model parameters are fixed.

[0061] S4. Construct a reverse design network model, including a generator network Generator for generating metasurface patterns, an image discriminator network Discriminator_A for determining the authenticity of patterns, and a spectral discriminator network Discriminator_B for determining the authenticity of spectra, and initialize the network parameters;

[0062] During the specific implementation, the reverse design network model is constructed, such as Figure 3 As shown in the figure, the generator network Generator is composed of deconvolution layers, which inputs a 51-dimensional absorbance vector Amp and generates a 64×64×1 matrix representing the metasurface pattern through deconvolution operations; the image discriminator network Discriminator_A is composed of convolution layers, which inputs the real image in the dataset and the image generated by the generator and the absorption spectrum into the discriminator, and the image features are converted into digital features, and the output discrimination probability is 0 or 1; the spectral discriminator Discriminator_B is composed of linear layers, which inputs the real spectrum and the predicted spectrum of the generated image, and outputs the discrimination probability 0 or 1.

[0063] S5. Send the training set in the reverse-engineered dataset into the generator network Generator, send the generated image and the real image into the image discriminator network Discriminator_A, send the spectrum predicted by the generated image and the real spectrum into the spectrum discriminator network Discriminator_B, and perform iterative cycle training.

[0064] In specific implementation, each round of iterative calculation process includes the following steps:

[0065] S5.1. Feed the absorbance vector of the reverse-engineered dataset into the generator network. The 51-dimensional spectral data A and the 20-dimensional Gaussian random noise z are concatenated using a concat operation. These are then fed into the generator's five deconvolutional network layers, where the number of nodes in the deconvolutional layers is 64, 32, 16, 8, and 1, respectively. Finally, the TanH activation function is used to constrain the output to the range [-1, 1] to generate a metamaterial structure image.

[0066] S5.2. The generated metamaterial structure image and the original metamaterial structure image corresponding to the absorbance vector are fed into the image discriminator network Discriminator_A and concatenated with the absorbance vector. After passing through five convolutional layers, the image features are converted into digital features. The image features are composed of five convolutional structures. Finally, the Sigmoid activation function is used for binary classification, and the output is 0 or 1. This is applied to the image classification to predict whether the input is a real pattern or a generated pattern. The loss of the image discriminator network Discriminator_A is defined as:

[0067]

[0068] Where m is the batch size, the superscript i represents the i-th data in the batch, the loss is the cross entropy between the image dataset x and the generated data G(A,z), and the discriminator output is between 0 and 1. In order to distinguish between real and fake pictures, D(x i) approaches 1, D(G(A i ,z i )) approaches 0, and this loss defines the similarity of the pattern between the image data and the generated image;

[0069] S5.3. The generated metamaterial structure image is fed into the forward prediction network. The predicted absorbance vector and the true absorbance vector are input into the spectral discriminator Discriminator_B. After passing through four linear layers, the sigmoid activation function is used for binary classification. The discriminant output is 0 or 1, and the predicted input is a true spectrum or a false spectrum. The loss of the spectral discriminator Discriminator_B is defined as:

[0070]

[0071] The loss is measured by the spectral data A and the predicted spectrum The cross entropy of the discriminator is between 0 and 1. In order to distinguish true and false spectra, D(A i ) approaches 1, Approaching 0, this loss defines the similarity between the spectral data and the predicted spectrum;

[0072] S5.4. Calculate the error using the image prediction value and the spectral prediction value, and perform a backpropagation algorithm to update the parameters of the real part generator network Generator, the image discriminator network Discriminator_A, and the spectral discriminator network Discriminator_B. The parameters of the generator are updated by backpropagation of the losses of the two discriminators:

[0073]

[0074] In order to make the generated image be able to deceive the discriminator well, in the game between the generator and the discriminator, D(G(A i ,z i ))and will approach 1. The reconstruction error of the spectral response is introduced, and λ is a parameter that balances the spectrum and image loss. In order to make the absorption spectrum of the reverse-designed graphic structure as similar as possible to the original graphic structure, rather than considering the similarity of the structure image, λ = 0.1 is selected.

[0075] Examples are given below to illustrate the present invention:

[0076] Example 1:

[0077] The training dataset for deep learning consists of 5800 MIM surface unit cell designs. Figure 4As shown, these designs are derived from six starting shape templates: open ring, cross, H-shape, bow tie, square, and donut. Full-wave simulations (under normal incidence TE polarization) were performed for each structure to generate corresponding 1001-point absorption spectra in the 120-160 THz frequency band. The response range in the training dataset can be expanded in future studies to enhance the predictive ability of the network. 80% of the data in the dataset was used as the training set, 10% as the validation set, and 10% as the test set. The image of each dimension in the dataset is a single-channel PNG image with a resolution of 64×64, and the corresponding label is a 51-dimensional vector Absorption.

[0078] Build the forward prediction network model according to step 3 and feed the forward prediction dataset into the forward prediction network model for iterative training. Using the Adam optimizer, dynamically adjust the learning rate. The initial learning rate is 0.01, and the learning rate is adjusted back to 0.1 every 250 training rounds. The training cycle is set to 1000.

[0079] Figure 5 The graph shows the decrease in the total loss value during 1000 training cycles, and the training set loss and test set loss are similar, eventually decreasing synchronously and converging to a lower value. The forward prediction network has high accuracy and generalization. After 1000 cycles of training, the loss on the validation set is 0.0052, which can obtain relatively accurate forward prediction results.

[0080] The parameters of the trained forward prediction network are fixed. According to step 4, the reverse design network model is constructed, including the generator network Generator for generating the metasurface pattern, the discriminator network Discriminator_A for judging the authenticity of the pattern, and the discriminator network Discriminator_B for judging the authenticity of the spectrum, and iterative loop training is performed.

[0081] The MAE of the absorption spectrum and the model accuracy are defined as:

[0082]

[0083] Randomly select individual cases from the test set and perform reverse design based on the absorption curve of the new pattern. The difference between the reverse design network design value and the simulation value is as follows: Figure 6 and Figure 7 As shown in Figure 1, the similarity between the designed and simulated values ​​of the reverse design network on the absorbance vector is extremely high. Specifically, the simulated values ​​of different supercell patterns are shown in Table 1. Compared to the traditional cDCGAN, the reverse design network has achieved a high level of accuracy, which can meet the needs of fast and accurate design.

[0084]

[0085] Table 1 Reverse design network simulation value table

[0086] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A hypersurface structure design method based on a joint discriminative generative adversarial network, characterized in that: The following steps are involved: S1. Designing a periodic arrangement of electromagnetic metasurface unit structures and related parameters, and inputting them into electromagnetic simulation software to obtain electromagnetic responses, constructing forward prediction datasets and reverse design datasets; both the forward prediction datasets and the reverse design datasets include generated electromagnetic response absorption spectrum data and metasurface patterns; S2. Split the forward prediction dataset and the reverse design dataset into training sets and test sets respectively, where the proportion of the training set is larger than the proportion of the test set; S3. Build a forward prediction network model, send the segmented forward prediction data set to the simulator network Simulator for forward prediction for training, and fix the forward prediction network model parameters after training is completed; S4. Construct a reverse design network model, including a generator network Generator for generating metasurface patterns, an image discriminator network Discriminator_A for determining the authenticity of patterns, and a spectral discriminator network Discriminator_B for determining the authenticity of spectra, and initialize the network parameters; S5. Send the training set in the reverse-engineered dataset to the generator network Generator, send the generated image and the real image to the image discriminator network Discriminator_A, send the spectrum predicted by the generated image and the real spectrum to the spectrum discriminator network Discriminator_B, and perform iterative cycle training; In S1, the steps of constructing a forward prediction dataset and a reverse design dataset include: Draw different metasurface unit patterns, construct corresponding electromagnetic metasurface units, and input them into electromagnetic simulation software to obtain the corresponding electromagnetic response; The encoding matrix of the metasurface unit pattern is used as the network input. In this process, the absorption rate Amp is obtained by comparing it with the reflectivity. and transmittance The relationship is calculated as follows: ； Through reflectivity and transmittance Calculate the absorption rate Amp, perform uniform frequency sampling on the absorption rate Amp, express it as a multi-dimensional vector, use it as the label of the data set, and construct a forward prediction data set; The absorption rate Amp is represented as a multi-dimensional vector and used as input at the same time, and its corresponding metasurface unit matrix is ​​used as a label to construct a reverse design dataset.

2. The method for designing a hypersurface structure based on a joint discriminative generative adversarial network according to claim 1, characterized in that: The electromagnetic metasurface unit includes an upper layer, a middle layer and a lower layer; the upper layer is divided into a coding pattern area; the middle layer dielectric substrate is made of F4B; and the lower layer is covered with full metal copper.

3. The method for designing a hypersurface structure based on a joint discriminative generative adversarial network according to claim 2, characterized in that: In the S3, the forward prediction network model is constructed, the simulator network Simulator adopts the ResNet18 structure, the optimizer of the network Simulator is set to Adam, and the learning rate is 1e-3; use = represents the network loss function, where is the true value, is the predicted value, The network is used as the dimension and the convergence of the model is improved through a dynamic adjustment mechanism. During the training process, the network's prediction of the absorption rate is gradually optimized. After the training is completed, the forward prediction network model parameters are fixed.

4. The method for designing a hypersurface structure based on a joint discriminative generative adversarial network according to claim 3, characterized in that: In the S4, a reverse design network model is constructed, in which the generator network Generator is composed of a deconvolution layer, receives an absorbance vector as input, and generates a matrix representing the metasurface pattern through a deconvolution operation; the image discriminator network Discriminator_A is composed of a convolution layer, and the real image in the data set and the image generated by the generator and the absorption spectrum are input into the discriminator, and the image features are converted into digital features to output the discrimination probability of the authenticity of the image; the spectral discriminator Discriminator_B is composed of a linear layer, which inputs the real spectrum and the predicted spectrum of the generated image, and outputs the discrimination probability of the authenticity of the spectrum.

5. The method for designing a hypersurface structure based on a joint discriminative generative adversarial network according to claim 4, characterized in that: In S2, the forward prediction data set input is a 64×64 encoding matrix, and the corresponding label is a 51-dimensional absorption rate Amp vector; the reverse design data set input is a 51-dimensional absorption rate Amp vector, and the corresponding label is a 64×64 encoding matrix.

6. The method for designing a hypersurface structure based on a joint discriminative generative adversarial network according to claim 5, characterized in that: In S5, each round of iterative calculation process includes the following steps: S5.

1. Feed the absorption coefficient vector of the reverse-engineered dataset into the generator network. This input vector is concatenated with random noise and processed through multiple deconvolution layers and activation functions to generate a metamaterial structure image. S5.

2. The generated metamaterial structure image and the original metamaterial structure image corresponding to the absorption rate vector are fed into the image discriminator network Discriminator_A. The input vector is concatenated with random noise and processed through multiple convolutional layers and activation functions to output a probability for determining the authenticity of the image. S5.

3. The generated metamaterial structure image is fed into the forward prediction network. The predicted absorbance vector and the true absorbance vector are input into the spectrum discriminator Discriminator_B. After processing through multiple linear layers and activation functions, the output is used to judge the probability of the spectrum authenticity. S5.

4. Calculate the error through the image prediction value and the spectral prediction value, and perform the back propagation algorithm to update the network parameters of the real part generator network Generator, the image discriminator network Discriminator_A, and the spectral discriminator network Discriminator_B.

7. The method for designing a hypersurface structure based on a joint discriminative generative adversarial network according to claim 6, characterized in that: In S5.2, the loss of the image discriminator network Discriminator_A is defined as: ； in is the batch size, superscript Indicates the first data, the loss is measured by the image dataset and the generated data The cross entropy of the discriminator is between 0 and 1. In order to distinguish true and false pictures, Approaching 1, Approaching 0, this loss defines the similarity of the pattern between the image data and the generated image.

8. The method for designing a hypersurface structure based on a joint discriminative generative adversarial network according to claim 7, characterized in that: In S5.3, the loss of the spectral discriminator Discriminator_B is defined as: ； The loss is measured by spectral data and predicted spectra The cross entropy of the discriminator is between 0 and 1. In order to distinguish true and false spectra, Approaching 1, Approaching 0, this loss defines the similarity between the spectral data and the predicted spectrum.

9. The method for designing a hypersurface structure based on a joint discriminative generative adversarial network according to claim 8, characterized in that: In S5.4, the parameters of the generator are updated by backpropagation through the two discriminator losses: ； In the game between the generator and the discriminator, and will approach 1, which introduces the reconstruction error of the spectral response, is a parameter that balances spectral and image loss, and is selected .

Images