Metasurface reverse design method of generative adversarial network based on anchor point network control

Abstract

The invention discloses a metasurface reverse design method of a generative adversarial network based on anchor point network control. The method comprises the following steps: collecting metasurface design parameters and spectral response data, preprocessing, and encoding the data into a format which can be processed by a deep learning model; and analyzing the training data by using a K-means clustering algorithm to generate a structured control vector. And constructing an AcGAN framework, and integrating the spectrum similarity measurement SOC into a loss function of a generator and a discriminator to enhance the spectrum matching precision. Through an adversarial training method, a generator is trained to generate a design conforming to target spectral characteristics, and meanwhile, a discriminator and an anchor network AncorNet are optimized to quickly predict and evaluate spectral response. Candidate designs are generated using the trained generator in combination with control vectors and random noise. And screening the candidate design to select the optimal design. According to the method, the spectrum regulation and control precision and the design efficiency of the metasurface design are effectively improved, and the challenge faced by a traditional method when a complex spectrum requirement is processed is solved.

Description

Technical Field

[0001] The present invention relates to a super surface design technology, and in particular to a super surface inverse design method based on a generative adversarial network controlled by an anchor point network. Background Art

[0002] In recent years, with the rapid development of photonic technology, metasurfaces have become a research hotspot in the field of optics and photonics due to their unique ability to manipulate electromagnetic wave properties at the nanoscale. Through structural design, metasurfaces can precisely control the amplitude, phase, and polarization state of light on a subwavelength scale, showing great application potential in optical devices, miniaturized sensors, imaging systems, and energy conversion devices. However, existing metasurface design methods mainly rely on empirical rules and finite element analysis. Although they can generate some structures with specific functions, existing methods are often unable to cope with complex spectral requirements. In particular, when designing metasurfaces with complex spectral characteristics, existing methods are difficult to simultaneously meet the requirements of spectral accuracy and computational efficiency.

[0003] Typical metal-insulator-metal (MIM) structure metasurfaces have a wide Lorentz absorption spectrum due to the plasmon resonance generated at the metal-dielectric interface. This structure has advantages in thermal radiation control and photothermal energy conversion applications. However, this type of metasurface has the problem of too broad spectral response during design, making it difficult to achieve fine spectral control. At the same time, hybrid dielectric metasurfaces generate Fano resonances through subwavelength cavity resonances, and their unique sharp spectral characteristics make them perform well in high-precision optical sensing. However, the design process of such structures is complex and highly sensitive to parameters, which easily leads to inconsistencies between design parameters and target spectra. Although existing inverse design methods can alleviate these problems to a certain extent, they are often limited to the application of traditional loss functions such as mean square error (MSE) and mean absolute error (MAE). These loss functions cannot capture key details in spectral characteristics, such as resonance peaks and specific absorption bands, resulting in obvious deficiencies in the generated metasurfaces in terms of accurate spectral matching.

[0004] In the face of the above challenges, existing technologies still have significant deficiencies in the fine-tuning of spectral characteristics and efficient design space exploration in metasurface design. Traditional design methods have difficulty in balancing accuracy and efficiency when dealing with complex spectral requirements, and conventional loss functions cannot effectively identify and optimize important details in the spectrum.

[0005] It should be noted that the information disclosed in the above background technology section is only used for understanding the background of the present application, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the invention

[0006] The main purpose of the present invention is to provide a metasurface inverse design method based on a generative adversarial network controlled by an anchor point network, so as to solve the problems of insufficient spectral control accuracy and low design efficiency existing in the prior art.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] In a first aspect of the present invention, a hypersurface inverse design method based on a generative adversarial network controlled by an anchor point network comprises the following steps:

[0009] S1. Determine the spectral characteristics required for the metasurface design and determine the target spectrum; collect and preprocess training data, including collecting metasurface design parameters and corresponding spectral response data, and preprocessing to adapt the deep learning model, encoding the metasurface design parameters into a format that can be processed by the deep learning model;

[0010] S2. Perform cluster analysis on the training data using the K-means clustering algorithm to generate structured control vectors associated with specific spectral categories;

[0011] S3. Construct an AcGAN framework consisting of a generator, a discriminator, and an anchor network AnchorNet. The spectral similarity metric SOC is introduced as a means to evaluate the similarity between the generated spectrum and the target spectrum and integrated into the loss function of the generator and the discriminator.

[0012] S4. Use adversarial training methods to train the generator to generate metasurface designs that meet the target spectral characteristics, and simultaneously train the discriminator and AnchorNet to quickly predict and evaluate the spectral response;

[0013] S5. Generate a series of candidate metasurface designs based on the control vector and random noise using the trained generator;

[0014] S6. Use evaluation tools to screen candidate designs and select the optimal design.

[0015] Furthermore, in step S1, the physical parameters of the metal-insulator-metal MIM and hybrid dielectric metasurface are encoded into a unified RGB image format, and appropriate data preprocessing is performed to ensure that the model can effectively learn and predict the spectral response;

[0016] Encode the MIM metasurface design parameters, including:

[0017] The design parameters of the MIM structure are encoded using the RGB image format, where each pixel represents a unit cell of a predetermined size;

[0018] Among them, the plasmon frequency of the metal resonator is encoded in the red channel of the RGB image, which determines the optical properties of the metal nanostructure;

[0019] Among them, the thickness of the dielectric layer is encoded in the blue channel of the RGB image, which adjusts the plasma resonance frequency and affects the absorption spectrum;

[0020] Among them, the green channel of the RGB image is not used in the encoding process and is fixed to zero;

[0021] Encode the hybrid dielectric metasurface design parameters, including:

[0022] The design parameters of the hybrid dielectric structure are encoded using an RGB image format, where each pixel represents a unit cell of a predetermined size;

[0023] Among them, the real refractive index of the dielectric resonator is encoded in the green channel of the RGB image as the core parameter affecting the optical performance of the dielectric;

[0024] The thickness of the dielectric layer is encoded in the blue channel of the RGB image, which affects the resonant behavior in the hybrid dielectric structure.

[0025] The red channel of the RGB image is not used during the encoding process and is fixed to zero.

[0026] Furthermore, in step S1, the preprocessing includes discretizing the spectral data, dividing the spectral data into a plurality of discrete points within a specific wavelength range, each discrete point corresponding to the spectral response intensity at a specific wavelength, and standardizing the discretized spectral data to scale parameters of different physical magnitudes to a uniform range.

[0027] Furthermore, step S2 specifically includes:

[0028] Spectral data clustering: Use a clustering algorithm to segment the training spectral data set and divide the data set into multiple clusters with unique spectral characteristics. Each cluster is represented by a centroid, which is the average characteristic of the spectral data in the cluster.

[0029] Determination of centroid: The centroid of each cluster is the representative of the spectral data in the cluster and is determined by optimizing the objective function of the clustering algorithm to ensure that the spectral characteristics of the data points in the cluster are highly similar to those of the centroid;

[0030] Calculation of spectral similarity metric SOC: Calculate the spectral similarity metric SOC to measure the similarity between a given spectrum and the centroid. The degree of overlap is directly measured by comparing the minimum and maximum values ​​of the spectrum at each wavelength. The closer the calculated SOC value is to zero, the higher the spectral similarity.

[0031] Generation of control vector: The calculated SOC values ​​are formed into a vector that describes the similarity information between the input spectrum and the predefined spectral categories, capturing the details and overall characteristics of the spectrum.

[0032] Construction of comprehensive control vector: The control vector is concatenated with the original spectral data to generate a comprehensive control vector, which is used as the input of the generator and the discriminator to guide the metasurface design process.

[0033] Furthermore, in step S3, the design of the generator specifically includes:

[0034] A design generator is configured to receive a control vector and a random noise vector as input, wherein the control vector contains original spectral data and cluster similarity information, and the random noise vector introduces randomness into the design space;

[0035] The network architecture of the generator converts the input vector into the spatial pattern of the hypersurface structure through a series of deconvolution layers, so that the generated hypersurface is similar to the real sample in structure and aligned with the target spectrum in spectral response;

[0036] The generator produces a metasurface structure as a candidate structure in the design space, which satisfies the set spectral property requirements;

[0037] Construct a comprehensive loss function for the generator, which consists of three parts: adversarial loss, spectral loss, and structural loss, to balance the generator's multiple optimization objectives when generating metasurface designs;

[0038] Among them, the adversarial loss is used to measure the probability that the hypersurface structure generated by the generator is judged as real by the discriminator, ensuring that the generator can generate high-quality and high-fidelity designs;

[0039] Among them, the spectral loss uses the spectral similarity metric (SOC) to evaluate the similarity between the generated metasurface spectral response and the target spectrum;

[0040] Among them, the structural loss uses the structural similarity index (SSIM) to measure the similarity between the generated hypersurface structure and the reference structure;

[0041] Different weight parameters are set to balance the impact of adversarial loss, spectral loss, and structural loss in the overall optimization process.

[0042] Furthermore, in step S3, the design of the discriminator specifically includes:

[0043] Design a discriminator to evaluate whether the generated hypersurface design conforms to the distribution of real data;

[0044] The discriminator network architecture analyzes the input hypersurface structure level by level through a series of convolutional layers to extract features that distinguish real from forged structures;

[0045] Construct the discriminator loss function, including adversarial loss, mismatch loss, and spectral loss, to accurately evaluate the generated design;

[0046] Among them, adversarial loss is used to evaluate the discriminator's ability to distinguish between real and generated designs, which is achieved by comparing the discriminator's scores on real samples and generated samples;

[0047] Among them, the mismatch loss enhances the discriminator's ability to detect inconsistencies between input conditions and generated designs, enabling the discriminator to identify designs that do not meet expectations;

[0048] Among them, the spectral loss uses the spectral similarity metric (SOC) to evaluate the similarity between the generated spectrum and the target spectrum.

[0049] Furthermore, in step S3, the design of the anchor network AnchorNet specifically includes:

[0050] Design AnchorNet to predict the spectral response of metasurface designs;

[0051] AnchorNet is built using the bottleneck ResNet architecture;

[0052] The loss function of AnchorNet is constructed based on the difference between the spectrum predicted by AnchorNet and the true spectrum obtained by electromagnetic simulation, with minimization as the optimization goal.

[0053] Furthermore, the network architecture of the generator includes:

[0054] The input layer receives a tensor that combines the control vector and random noise;

[0055] Multi-layer convolution and deconvolution operations are used to generate three-dimensional hypersurface images from high-dimensional space, with batch normalization and Leaky ReLU activation function after each layer;

[0056] The output layer uses the Tanh activation function to make the pixel value of the generated image in the range of [-1, 1], which is consistent with the standardization of the input data;

[0057] The network architecture of the discriminator includes:

[0058] an input layer that receives the metasurface image for evaluating structural and spectral properties;

[0059] Multi-layer convolutional layers are used to extract input image features, using batch normalization and Leaky ReLU activation functions;

[0060] The fully connected layer maps the features extracted by the convolutional layer into a one-dimensional vector and outputs a probability value through the Sigmoid activation function, which represents the image authenticity score;

[0061] The network framework of the anchor network AnchorNet includes:

[0062] An input layer that receives the metasurface image for spectral response prediction;

[0063] The residual bottleneck block is used for deep feature extraction, retaining important information through convolutional layers, batch normalization and ReLU activation functions, and skip connections; multiple bottleneck blocks perform convolution operations of different scales to extract and retain image features;

[0064] The output layer outputs a one-dimensional spectral vector. The extracted features are mapped to the spectral response space through the convolution layer and the fully connected layer. The activation function is selected according to the requirements of the spectrum to make the output spectrum value range meet the expectations.

[0065] Furthermore, in step S4, the training process of AcGAN includes:

[0066] The training data includes reference metasurface, target spectral samples, and control vectors;

[0067] Initialization operation, parameter initialization of the generator and discriminator, including random initialization of weights;

[0068] AnchorNet pre-training, which is pre-trained to predict the spectral characteristics of metasurface designs;

[0069] Training iteration: shuffling data, sampling noise samples and control vectors, and obtaining sample pairs for each training round;

[0070] Discriminator update, which performs multiple update steps of the discriminator, including generating the hypersurface, evaluating the output of the reference and generated hypersurfaces, and the output of the mismatched data;

[0071] Discriminator loss calculation: calculate the discriminator loss according to a specific formula and update the discriminator parameters accordingly;

[0072] Generator update, generating the hypersurface and calculating the generator loss, followed by updating the generator parameters;

[0073] After training is completed, the trained discriminator and generator models are output.

[0074] In a second aspect of the present invention, a computer program product includes a computer program, which, when executed by a processor, implements the hypersurface inverse design method based on a generative adversarial network controlled by an anchor point network.

[0075] The present invention has the following beneficial effects:

[0076] The present invention proposes a hypersurface inverse design method based on Generative Adversarial Network (AcGAN), provides an innovative design framework, and significantly improves the spectral matching accuracy and design efficiency of the hypersurface design by introducing an innovative spectral similarity metric (SOC) and a structured control vector. The hypersurface inverse design method of the present invention can accurately capture spectral features, improve design efficiency, and is applicable to a variety of hypersurface types.

[0077] First, based on the existing mean square error (MSE) and mean absolute error (MAE), the present invention innovatively proposes to introduce a spectral similarity metric (SOC) as part of the loss function evaluation to accurately measure the degree of overlap between the generated spectrum and the target spectrum. By calculating the ratio of the minimum and maximum values ​​between the spectra, SOC can better capture key features in the spectrum such as resonance peaks and specific absorption bands, greatly improving the accuracy of metasurface spectral regulation. This innovation directly solves the deficiency of traditional loss functions that cannot capture spectral details, ensuring that the generated metasurface is more accurate in spectral matching.

[0078] Secondly, the present invention introduces structured control vectors into the design framework, segments and clusters the training spectral data through the K-means clustering algorithm, generates control vectors associated with specific spectral categories, and combines them with the original spectral data to input the generator and discriminator. This two-layer data processing method not only retains the detailed information of the spectrum, but also enhances the exploration ability of the model in the design space, avoiding the error amplification problem that may occur in traditional methods when dealing with complex spectral requirements. This innovation effectively solves the problem of insufficient accuracy and efficiency of existing design methods when dealing with complex spectral characteristics, so that the generated metasurface design can meet stringent spectral requirements in a variety of application scenarios.

[0079] In addition, the present invention also uses the anchor network (AnchorNet) module in the AcGAN framework to quickly predict the spectral response of the metasurface design, reducing the calculation time of the design iteration and improving the overall design efficiency. The anchor network not only provides accurate spectral response prediction in the adversarial generation process, but also prevents model overfitting through the early stopping mechanism, ensuring the wide applicability and efficiency of the generated metasurface design. These innovations comprehensively enhance the design capabilities of the metasurface and provide more powerful technical support for advanced applications in the field of photonics.

[0080] Through the innovative design method of the present invention, the shortcomings of the existing technology in terms of fine control of spectral characteristics and efficient design exploration are overcome, providing a solid foundation for the development of next-generation optical devices and sensing technologies.

[0081] Other beneficial effects of the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 This is a framework diagram of the AcGAN model according to an embodiment of the present invention.

[0083] Figure 2 This is a detailed network architecture diagram of the three core components of the generator, discriminator and anchor network AnchorNet in the AcGAN framework of an embodiment of the present invention. DETAILED DESCRIPTION

[0084] The following is a detailed description of the embodiments of the present invention. It should be emphasized that the following description is only exemplary and is not intended to limit the scope and application of the present invention.

[0085] The embodiment of the present invention provides a hypersurface inverse design method based on a generative adversarial network controlled by an anchor point network, which mainly includes the following steps:

[0086] S1. Determine the spectral characteristics required for the metasurface design and determine the target spectrum; collect and preprocess training data, including collecting metasurface design parameters and corresponding spectral response data, and preprocessing to adapt the deep learning model, encoding the metasurface design parameters into a format that can be processed by the deep learning model;

[0087] S2. Perform cluster analysis on the training data using the K-means clustering algorithm to generate structured control vectors associated with specific spectral categories;

[0088] S3. Construct AcGAN framework: Construct AcGAN framework including generator, discriminator and anchor network AnchorNet (see Figure 1 and Figure 2 ), where the spectral similarity metric SOC is introduced as a means to evaluate the similarity between the generated spectrum and the target spectrum and integrated into the loss function of the generator and the discriminator;

[0089] S4. Use adversarial training methods to train the generator to generate metasurface designs that meet the target spectral characteristics, and simultaneously train the discriminator and AnchorNet to quickly predict and evaluate the spectral response;

[0090] S5. Generate a series of candidate metasurface designs based on the control vector and random noise using the trained generator;

[0091] S6. Use evaluation tools such as SOC and other evaluation criteria such as spectral performance and structural reliability to evaluate and screen candidate designs and select the optimal design.

[0092] The present invention provides an innovative inverse design method for supersurfaces based on a generative adversarial network controlled by an anchor network. The method significantly improves the accuracy and efficiency of supersurface design through several key innovations. First, the present invention introduces the spectral similarity metric SOC as a means to evaluate the similarity between the generated spectrum and the target spectrum, and integrates it into the loss function of the generator and the discriminator. As a result, the loss function can accurately measure the similarity between the generated spectrum and the target spectrum, effectively capture key details such as the resonance peak and specific absorption band in the spectrum, thereby solving the shortcomings of the traditional loss function in fine spectral regulation. Secondly, the present invention introduces a structured control vector, and uses the K-means clustering algorithm to effectively segment and cluster the training data, thereby enhancing the model's ability to explore the design space, while avoiding the problem of error amplification under complex spectral requirements. In addition, the present invention constructs the anchor network (AnchorNet) module in the AcGAN framework, which realizes the rapid prediction of the spectral response of the supersurface design, greatly reduces the calculation time of the design iteration, and improves the efficiency of the design process. The present invention not only improves the spectral matching accuracy of metasurface design, but is also applicable to a variety of metasurface types, providing strong technical support for advanced applications in the field of photonics and laying a solid foundation for the development of next-generation optical devices and sensing technologies.

[0093] The specific embodiments of the present invention are further described below.

[0094] 1. Encoding of metasurface design parameters and data preprocessing

[0095] First, during the metasurface design process, the design parameters of different structures need to be converted into a format that can be processed by the deep learning model for subsequent model training and optimization. The main goal of this step is to encode the physical parameters of metal-insulator-metal (MIM) structures and hybrid dielectric metasurfaces into a unified RGB image format and perform appropriate data preprocessing to ensure that the model can effectively learn and predict the spectral response.

[0096] MIM metasurface coding:

[0097] The MIM structure has a wide Lorentz absorption spectrum due to the plasma resonance generated at the metal-dielectric interface. This type of structure is often used in fields such as thermal radiation control and photothermal energy conversion. In order to effectively represent the design parameters of the MIM structure, a 3×64×64 RGB image format is used for encoding, where each pixel represents a 50 nanometers × 50 nanometers unit cell. The specific mapping is as follows:

[0098] Red channel (R): Encoded metal resonator plasma frequency ω p The plasmon frequency is a key parameter that determines the optical properties of metal nanostructures, affecting both the resonance wavelength and the absorption characteristics.

[0099] Blue channel (B): The thickness of the coding dielectric layer, d, in nanometers. The thickness of the dielectric layer adjusts the plasma resonance frequency and directly affects the absorption spectrum.

[0100] Green channel (G): This channel is not used and is fixed to 0 to ensure simplicity and uniformity of encoding.

[0101] Hybrid Dielectric Metasurface Coding:

[0102] The hybrid dielectric metasurface generates Fano resonance through subwavelength cavity resonance, and its sharp spectral characteristics excel in high-precision optical sensing. In order to encode the design parameters of this structure into a form that can be processed by the deep learning model, a 3×64×64 RGB image format is used, where each pixel represents a unit cell of 120 nanometers by 120 nanometers. The specific mapping is as follows:

[0103] Green channel (G): The real refractive index n of the coded dielectric resonator. The refractive index is the core parameter affecting the optical properties of dielectrics and directly affects the characteristics of Fano resonance.

[0104] Blue channel (B): The thickness d of the coding dielectric layer, in nanometers, is the same as in the MIM structure. This parameter is crucial to the resonant behavior in the hybrid dielectric structure.

[0105] Red channel (R): This channel is not used and is fixed at 0.

[0106] Spectral discretization and normalization:

[0107] To ensure that the model can process and learn the spectral response data, the spectral data in the range of 4-12 microns is discretized into 800 discrete points. Each discrete point represents the spectral response intensity at a specific wavelength. This discretization process can effectively capture the details in the spectral characteristics while simplifying the data representation, making it suitable for the input of the deep learning model.

[0108] Next, the encoded image data is normalized. The purpose of normalization is to scale parameters of different physical magnitudes to a uniform range (usually [0, 1] or [-1, 1]) to eliminate the magnitude differences between different physical parameters and ensure that the deep learning model can better capture the relationship between parameters. This step greatly improves the stability and efficiency of model training.

[0109] The comprehensive significance of encoding and data preprocessing:

[0110] Through this encoding and data preprocessing method, a unified input format is provided for the deep learning model, so that different types of metasurface designs can be optimized and analyzed under the same framework. The encoding method of RGB images enables the model to capture the spatial and spectral correlations of design parameters using the convolution operation in the convolutional neural network. At the same time, the spectral discretization and standardization processing ensure the consistency and stability of the model input data, thereby improving the prediction accuracy and generalization ability of the model. In practical applications, this method is not only applicable to the design of MIM and hybrid dielectric metasurfaces, but can also be extended to other types of nanostructure designs as needed. By adjusting the pixel scale of the encoding and introducing more physical parameters (such as dielectric constant, anisotropic properties, etc.), the applicability and flexibility of the model can be further improved, enabling it to handle more complex design tasks and a wider range of spectral requirements. This preprocessing and encoding method lays the foundation for the entire inverse design framework, ensuring that the model has sufficient accuracy and efficiency when dealing with complex spectral responses and design tasks to meet diverse application needs.

[0111] 2. Clustering of spectral data and generation of control vectors

[0112] In the process of metasurface design, due to the complexity and diversity of spectral responses, it is often difficult to capture key features in the spectrum, such as resonance peaks or specific absorption bands, by relying solely on traditional loss functions (such as mean square error (MSE) or mean absolute error (MAE)) to optimize design parameters. Therefore, this paper introduces a new method based on spectral data clustering and control vector generation to better characterize and utilize important information in spectral data.

[0113] Clustering of spectral data:

[0114] First, a set of training spectral data sets S =

[0115] {s1,s2,…,s n}, where each s i Represents a data point with a unique spectral response. In order to effectively process these spectral data and extract representative spectral features, the K-means clustering algorithm is used to segment and cluster the spectral data.

[0116] The goal of the K-means clustering algorithm is to divide the spectral dataset S into k different clusters C =

[0117] {C1,C2,…,C k}, and optimize the clustering results by minimizing the following objective function:

[0118]

[0119] Among them, ci represents the i-th cluster C i The centroid (i.e., cluster center) of , which represents the average characteristics of the spectral data in the cluster.

[0120] Through K-means clustering, each cluster C i The spectral data points in the cluster are all related to the centroid c of the cluster. i As close as possible in the spectral feature space, thus ensuring that the spectral data in each cluster have highly similar spectral characteristics. After clustering, k centroids are obtained, each of which can be regarded as a representative of a certain type of spectral feature.

[0121] Calculation of SOC (spectral similarity measure):

[0122] In obtaining the centroid c i Finally, a new spectral similarity metric (SOC) is proposed to measure the similarity between a given spectrum and these centroids. SOC is calculated as follows:

[0123]

[0124] Where s represents a specific spectral response, c i represents the i-th centroid. The closer the SOC value is to zero, the closer the SOC value is to zero. i The higher the spectral similarity.

[0125] SOC directly measures the overlap between two spectra by comparing their minimum and maximum values ​​at each wavelength. Unlike traditional loss functions, SOC is better able to capture key features in the spectrum, especially those resonance peaks or absorption bands that may only account for a small part of the entire spectrum but are crucial.

[0126] Control vector generation:

[0127] Once the spectrum s and each centroid c are calculated i The SOC value between is represented as a k-dimensional vector:

[0128] v=[SOC(s,c1),SOC(s,c2),…,SOC(s,c k )] (3)

[0129] The vector v quantitatively describes the similarity information between the input spectrum s and k predefined spectrum categories (i.e., centroids). The similarity vector not only contains the detailed information of the spectrum, but also captures the overall characteristics of the spectrum by clustering.

[0130] Next, the vector v is concatenated with the original spectral data s to generate a comprehensive control vector u:

[0131] u=[s;v] (4)

[0132] The control vector u combines the original information of the spectrum and the similarity information after clustering, and serves as the input of the generator and discriminator to guide the subsequent metasurface design process.

[0133] Application of control vectors in generators and discriminators:

[0134] In the generator, the control vector u is used as input to generate a metasurface structure that meets specific spectral requirements. By combining the original spectral data with the clustering information, the generator can better capture the important features in the spectrum and generate a design that better meets the target requirements.

[0135] In the discriminator, the control vector u is used to evaluate whether the generated metasurface design meets the expected spectral response. Since the control vector contains clustering information, the discriminator can more accurately identify the subtle differences between the generated design and the target spectrum, thereby improving the design quality.

[0136] Comprehensive significance:

[0137] By introducing the clustering of spectral data and the generation of control vectors, the present invention significantly improves the spectral matching accuracy and design efficiency in the metasurface design process. As a new spectral similarity measure, SOC effectively solves the problem that traditional loss functions cannot capture spectral details. Through the generation and application of control vectors, the entire design process can explore the design space more effectively and ensure that the generated metasurface structure has high accuracy and consistency in spectral response. This method is not only suitable for current metasurface design needs, but can also be extended to the design of other nanostructures with complex spectral characteristics by adjusting the clustering algorithm and the generation method of control vectors, providing support for a wider range of photonics applications. 3. Hypersurface design generation based on AcGAN

[0138] In order to achieve accurate matching of complex spectral responses in metasurface design, a generation method called Anchor-controlled Generative Adversarial Network (AcGAN) is proposed. The AcGAN framework is an innovative network structure proposed by this invention, which integrates four core components: controller, generator, discriminator, and anchor network (AnchorNet). Through adversarial training, the design parameters are optimized to generate a metasurface structure that meets specific spectral requirements.

[0139] Overview of AcGAN framework:

[0140] The core of the AcGAN framework is to continuously optimize the generated metasurface design through adversarial training of the generator and the discriminator. The generator is responsible for generating a metasurface structure with realistic spectral characteristics from the input control vector u and random noise vector z, while the discriminator is responsible for evaluating whether these generated designs meet the expected spectral response and structural requirements. Figure 1 shown.

[0141] Generator design and input:

[0142] The generator receives as input a control vector u and a random noise vector z. The control vector u contains the original spectral data and cluster similarity information, while the noise vector z introduces randomness in the design space, allowing the generator to explore diverse design options. The generator converts these inputs into spatial patterns of the hypersurface structure through a series of deconvolutional layers. The main task of the generator is to ensure that the generated hypersurface is not only similar to the real sample in structure, but also aligned with the target spectrum in spectral response.

[0143] Generator output and loss function:

[0144] The output of the generator is a hypersurface structure M g , which represents the candidate structures generated in the design space. In order to quantify the performance of the generator, a comprehensive loss function L is proposed G , consists of the following three parts:

[0145]

[0146] Fighting losses Measures whether the generator has successfully deceived the discriminator. The specific formula is:

[0147]

[0148] Where D(G(z,u),u) is the identification score given by the discriminator, which indicates the probability that the generated design is judged to be true.

[0149] Spectral loss L spectral :The spectral similarity metric (SOC) is proposed to evaluate the spectral response of the generated metasurface g With the target spectrum t The specific formula is:

[0150] L spectral =SOC(s g ,s t ) (7)

[0151] Structural loss L structural:It is proposed to use the structural similarity index (SSIM) to measure the generated hypersurface structure M g With reference structure M r The specific formula is:

[0152] L structural =SSIM(M g ,M r ) (8)

[0153] The parameters γ, α, and β in these loss functions are used to balance the weights of adversarial, spectral, and structural losses in the overall optimization process. The generator minimizes the comprehensive loss L G , and continuously improves its output so that the generated metasurface designs are both spectrally compliant and structurally sound.

[0154] Design and function of the discriminator:

[0155] The discriminator is one of the key components in the AcGAN framework. It is responsible for evaluating the generated hypersurface design to determine whether it conforms to the distribution of real data. The discriminator analyzes the input hypersurface structure step by step through a series of convolutional layers to extract features that can distinguish between real and fake structures.

[0156] The loss function L of the discriminator D It contains several parts to ensure that it can accurately evaluate the generated design:

[0157]

[0158] Fighting losses Evaluate the discriminator's ability to distinguish between real and generated designs. The specific formula is:

[0159]

[0160] Among them, D(M r ,s t ) represents the score of the discriminator for the real sample, and D(G(z,u)) represents the score of the generated sample.

[0161] Mismatch loss L mismatch : It is proposed to enhance the discriminator's ability to detect inconsistencies between input conditions and generated designs. The specific formula is:

[0162]

[0163] Here, u′ is a control vector that does not match z, ensuring that the discriminator can identify designs that do not meet expectations.

[0164] Spectral loss Lspectral : As in the generator, the discriminator evaluates the generated spectrum s by SOC g With the target spectrum t The similarity between .

[0165] Introduction of AnchorNet:

[0166] AnchorNet is a module proposed and designed in the AcGAN framework for fast prediction of spectral responses. It uses a bottleneck ResNet architecture to reduce computational complexity and optimize the accuracy of spectral prediction. The main goal of AnchorNet is to minimize the predicted spectrum. The gap between the true spectrum s is optimized by the following loss function:

[0167]

[0168] in, is the spectrum predicted by AnchorNet, and s is the real spectrum from electromagnetic simulation.

[0169] AcGAN training process:

[0170] In order to enable the generator and discriminator to generate and evaluate high-quality hypersurface designs, the following training process is proposed:

[0171] In the proposed AcGAN framework, the training process of hypersurface design is crucial. To describe the training process more clearly, the following is a Chinese description based on pseudocode, outlining the steps of the entire training process.

[0172] Algorithm 1: AcGAN training process for metasurface design Input: Training pair in is the reference metasurface, is the target spectrum sample, u i is the control vector.

[0173] Parameters: number of training epochs E, batch size B, learning rate η, spectral and structural loss weights α, β, adversarial and mismatch loss weights γ.

[0174] 1. Initialize the parameters θ of the generator G G and the parameters θ of the discriminator D D , randomly initialize the weights.

[0175] 2. Pre-training AnchorNet to predict spectral properties from metasurface designs;

[0176] 3. For each training round epoch = 1 to E, do the following:

[0177] 4. Shuffle the training data;

[0178] 5. From the noise distribution p Z (z) to obtain m noise samples {z (1) ,…,z (m)}; From the control vector distribution p U (u) to obtain m control vectors {u (1) ,…,u (m)}; From the data distribution p data (M r ,s t ) to obtain m sample pairs

[0179]

[0180] 6. For each batch {z,u,x} of size B, do the following:

[0181] Update the discriminator

[0182] 7. Perform k discriminator update steps:

[0183] 8.M g ←G(z,u) / / Generate a hypersurface with condition u;

[0184] 9.D real ←D(x,u) / / Output of the discriminator to the reference hypersurface;

[0185] 10.D fake ←D(M g ,u) / / Output of the discriminator to generate the hypersurface;

[0186] 11.D mismatch ←D(M g ,u′) / / Output of the discriminator for mismatched data;

[0187] 12. Calculate the loss L of the discriminator D :

[0188]

[0189] 13. Update the discriminator parameters θ D :

[0190]

[0191] 14. End the discriminator update step.

[0192] Update Generator

[0193] 15.M g←G(z,u) / / Generate a hypersurface with condition u;

[0194] 16.D fake ←D(M g ,u) / / Output of the discriminator to generate the hypersurface;

[0195] 17. Calculate the generator loss L G :

[0196]

[0197] 18. Update the generator parameters θ G :

[0198]

[0199] Output: The trained discriminator D and generator G models.

[0200] Through this iterative training, the AcGAN framework is able to gradually generate high-quality metasurface designs, so that the final generated structure is not only visually similar to the real structure, but also highly matches the target spectrum in terms of spectral response.

[0201] Comprehensive significance:

[0202] By proposing and applying the AcGAN framework, this paper achieves efficient generation and optimization of metasurface design. The generator can explore a wide range of design spaces, the discriminator provides strict design quality control, and the AnchorNet ensures the accuracy of spectral prediction. This integrated generation method makes it possible to design metasurfaces with complex spectral requirements, greatly improving the design efficiency and accuracy.

[0203] 4. Network Architecture

[0204] In the AcGAN framework, the generator, discriminator, and anchor network (AnchorNet) are three core components, each of which undertakes specific tasks and achieves these tasks through a carefully designed network structure. The following is a detailed description of the network architecture of these three components: Figure 2 As shown. Generator

[0205] The main task of the generator is to receive the control vector u and the random noise vector z and generate a realistic hypersurface structure M through a series of deconvolution layers. g The network architecture of the generator is shown in Figure (a), which includes several key parts:

[0206] Input layer: The input of the generator is a tensor of size 2000×1×1, which contains a combination of the control vector u and the random noise z.

[0207] Convolutional and deconvolutional layers: The generator uses multiple layers of convolutional and deconvolution operations to gradually reduce the dimensionality from the high-dimensional space and generate the final three-dimensional hypersurface image. Each layer is accompanied by batch normalization (BatchNorm) and Leaky ReLU activation functions to stabilize the training process and improve the quality of the generated image.

[0208] Output layer: The output of the generator is a tensor of size 3×64×64, representing the generated hypersurface structure M g The output layer uses the Tanh activation function to make the pixel values ​​of the generated image in the range of [-1,1], which is consistent with the standardization of the input data.

[0209] The task of the discriminator is to receive the generated hypersurface structure M g and the reference metasurface structure M r , evaluate them and judge their authenticity. The architecture of the discriminator is shown in Figure (b), which has the following characteristics:

[0210] Input layer: The input to the discriminator is a 3×64×64 metasurface image (generated or reference) for evaluation of structural and spectral properties.

[0211] Convolutional layer: The discriminator gradually extracts the features of the input image through multiple layers of convolution. The convolution operation is accompanied by batch normalization and Leaky ReLU activation function to ensure that the discriminator can extract important spatial features.

[0212] Fully connected layer: After the convolution operation, the discriminator maps the extracted features to a one-dimensional vector through the fully connected layer, and outputs a probability value D between [0,1] through the Sigmoid activation function, which represents the authenticity score of the input image.

[0213] AnchorNet

[0214] The main function of the anchor network is to predict the spectral response of the generated metasurface design. g AnchorNet achieves fast and accurate spectral prediction through a deep residual network architecture. Its architecture is shown in Figure (c), which includes the following key parts:

[0215] Input layer: The input of AnchorNet is a 3×64×64 hypersurface image (generated or reference), which is the same as the input of the discriminator.

[0216] Residual Bottleneck Block: AnchorNet uses residual bottleneck blocks to process input data. These bottleneck blocks extract deep features by using convolutional layers, batch normalization, and ReLU activation functions, and retain important information in the image through skip connections (ShortCut).

[0217] The network contains multiple bottleneck blocks, which perform convolution operations of different scales to extract features.

[0218] Output layer: The output of AnchorNet is a one-dimensional spectral vector s g , the extracted features are mapped to the spectral response space through the convolution layer and the fully connected layer. The activation function of the output layer uses different activation functions (such as Tanh or Sigmoid) according to the requirements of the specific spectrum to ensure that the numerical range of the output spectrum meets the expectations.

[0219] These three networks work together to achieve an efficient and accurate generation process of hypersurface design in the AcGAN framework. The generator continuously improves its ability to generate hypersurface structures through adversarial training with the discriminator and anchor network. The discriminator provides feedback to improve the performance of the generator by evaluating the authenticity and spectral consistency of the generated structure. The anchor network provides additional optimization guidance to the generator by quickly and accurately predicting the spectral response.

[0220] Through these carefully designed network architectures, AcGAN is able to effectively cope with complex metasurface design tasks, ensuring that the generated designs have high-quality spectral properties and structural rationality for practical applications.

[0221] 5. Design optimization and iterative improvement

[0222] In the AcGAN framework, after multiple rounds of iterative training, the generator outputs a series of candidate hypersurface designs. These designs need to be further verified through design exploration and optimization to ensure their performance in practical applications. Section 5 will describe in detail how to screen, test, and finally optimize these candidate designs.

[0223] Design exploration and candidate generation

[0224] After completing multiple rounds of training, the generator not only outputs a single design, but also generates a set of candidate designs with diverse characteristics. These designs reflect the various possibilities captured by the generator when exploring the design space. Specifically, they include:

[0225] Designs with different structural parameters: Due to the numerous possible combinations of the control vector u and the random noise vector z, the generator can output metasurface designs with a variety of structural parameter combinations to meet different spectral requirements.

[0226] Diversified spectral response: The designs generated by the generator may have different characteristics in spectral response, such as broadband absorption, narrowband selective absorption, high reflection in specific bands, etc. These spectral characteristics enable candidate designs to cover a wider range of application scenarios.

[0227] Screening of candidate designs

[0228] Selecting the best solution from the candidate designs output by the generator is a key step to achieve efficient application. This process usually includes the following steps:

[0229] Spectral performance screening: The spectral response of each candidate design is accurately calculated and compared with the target spectrum using electromagnetic simulation tools (such as Lumerical FDTD). Designs that perform best on key spectral parameters are selected.

[0230] Structural reliability assessment: The physical structure of the candidate design is tested for stability, assessing its feasibility in actual manufacturing and its structural integrity under various environmental conditions. This includes simulating thermal effects, mechanical stress, material aging and other factors that may affect the design performance.

[0231] Practical application testing The screened designs need to be tested in the actual application environment to verify whether their performance meets expectations. Practical application testing includes:

[0232] Physical manufacturing and prototype testing: The candidate design is manufactured into actual samples and prototypes are tested under laboratory conditions. The actual performance of the design is verified by measuring actual parameters such as spectral response, reflectivity, and transmittance.

[0233] Environmental adaptability testing: Test the prototype design under different environmental conditions (such as high temperature, humidity, and mechanical stress) to ensure its long-term stability and reliability in different application scenarios.

[0234] The final choice for design optimization

[0235] After actual application testing, the final design selection is based not only on the precise matching of spectral response, but also on comprehensive consideration of manufacturing cost, material selection, structural complexity, etc. The final selected design will be further optimized to meet the specific application requirements. The optimization process may include:

[0236] Material optimization: Optimize material selection based on actual manufacturing needs to reduce costs while maintaining high performance.

[0237] Process improvement: Based on the limitations of the manufacturing process, the design is appropriately simplified or adjusted to improve production efficiency and yield.

[0238] Feedback and Continuous Improvement

[0239] The final selected design may still need to be fine-tuned and improved in the actual application process. This process feeds the generated design back to the AcGAN training framework for secondary optimization, so that the generator can output a design solution that better meets the application requirements in subsequent design tasks.

[0240] Conclusion: Through systematic design exploration and optimization verification, the AcGAN framework not only achieves efficient metasurface design generation, but also ensures that the generated designs perform well in practical applications. This process provides strong support for future complex photonics and optical applications.

[0241] In summary, this paper proposes a metasurface inverse design method based on anchor network-controlled generative adversarial network (AcGAN), which improves the spectral matching accuracy and design efficiency of metasurface design through the following innovations:

[0242] A spectral similarity metric (SOC) was introduced. This metric is used to evaluate the similarity between the generated spectrum and the target spectrum. By calculating the ratio of the minimum and maximum values ​​of the two at each wavelength, it can effectively capture the key features in the spectrum, such as resonance peaks and specific absorption bands, thereby improving the accuracy of spectral regulation.

[0243] A structured control vector is adopted to analyze the training spectral data through the K-means clustering algorithm to generate control vectors associated with specific spectral categories. These vectors are combined with the original spectral data and input into the generator and discriminator, which enhances the model's ability to explore the design space and retains the detailed information of the spectrum.

[0244] By utilizing the AnchorNet module in the AcGAN framework, the spectral response of the metasurface design can be quickly predicted, the computational time of design iterations can be reduced, and the model overfitting can be prevented through an early stopping mechanism, thus improving the wide applicability and efficiency of the design.

[0245] By comprehensively applying the above-mentioned innovations, the metasurface inverse design method of the present invention not only improves the spectral matching accuracy of the design, but also significantly improves the design efficiency. It is applicable to a variety of metasurface types and provides a solid technical foundation for advanced applications in the field of photonics and the development of next-generation optical devices and sensing technologies.

[0246] An embodiment of the present invention further provides a storage medium for storing a computer program, which at least performs the above method when executed.

[0247] An embodiment of the present invention further provides a control device, comprising a processor and a storage medium for storing a computer program; wherein the processor is configured to execute at least the method described above when executing the computer program.

[0248] An embodiment of the present invention further provides a processor, wherein the processor executes a computer program and at least executes the method described above.

[0249] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a ferromagnetic random access memory (Flash Memory), a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory can be a disk memory or a tape memory. The storage medium described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0250] In the several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0251] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0252] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0253] A person skilled in the art can understand that: all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, etc. Various media that can store program codes.

[0254] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0255] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0256] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0257] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0258] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A hypersurface inverse design method based on a generative adversarial network controlled by an anchor network, characterized in that: The following steps are involved: S1. Determine the spectral characteristics required for the metasurface design and determine the target spectrum; Collecting and preprocessing training data, including collecting metasurface design parameters and corresponding spectral response data, and preprocessing to adapt the deep learning model, encoding the metasurface design parameters into a format that can be processed by the deep learning model; S2. Perform cluster analysis on the training data using the K-means clustering algorithm to generate structured control vectors associated with specific spectral categories; S3. Construct an AcGAN framework consisting of a generator, a discriminator, and an anchor network AnchorNet. The spectral similarity metric SOC is introduced as a means to evaluate the similarity between the generated spectrum and the target spectrum and integrated into the loss function of the generator and the discriminator. S4. Use adversarial training methods to train the generator to generate metasurface designs that meet the target spectral characteristics, and simultaneously train the discriminator and AnchorNet to quickly predict and evaluate the spectral response; S5. Generate a series of candidate metasurface designs based on the control vector and random noise using the trained generator; S6. Use evaluation tools to screen candidate designs and select the optimal design.

2. The hypersurface inverse design method based on the generative adversarial network controlled by the anchor network as claimed in claim 1, characterized in that: In step S1, the physical parameters of the metal-insulator-metal MIM and the hybrid dielectric metasurface are encoded into a unified RGB image format and data preprocessing is performed; Encode the MIM metasurface design parameters, including: The design parameters of the MIM structure are encoded using the RGB image format, where each pixel represents a unit cell of a predetermined size; Among them, the plasmon frequency of the metal resonator is encoded in the red channel of the RGB image, which determines the optical properties of the metal nanostructure; Among them, the thickness of the dielectric layer is encoded in the blue channel of the RGB image, which adjusts the plasma resonance frequency and affects the absorption spectrum; Among them, the green channel of the RGB image is not used in the encoding process and is fixed to zero; Encode the hybrid dielectric metasurface design parameters, including: The design parameters of the hybrid dielectric structure are encoded using an RGB image format, where each pixel represents a unit cell of a predetermined size; Among them, the real refractive index of the dielectric resonator is encoded in the green channel of the RGB image as the core parameter affecting the optical performance of the dielectric; The thickness of the dielectric layer is encoded in the blue channel of the RGB image, which affects the resonant behavior in the hybrid dielectric structure. The red channel of the RGB image is not used during the encoding process and is fixed to zero.

3. The hypersurface inverse design method based on the anchor point network controlled generative adversarial network according to any one of claims 1 to 2, characterized in that: In step S1, the preprocessing includes discretizing the spectral data, dividing the spectral data into multiple discrete points within a specific wavelength range, each discrete point corresponding to the spectral response intensity at a specific wavelength, and standardizing the discretized spectral data to scale parameters of different physical magnitudes to a uniform range.

4. The hypersurface inverse design method based on the anchor point network controlled generative adversarial network according to any one of claims 1 to 3, characterized in that: Step S2 specifically includes: Spectral data clustering: Use a clustering algorithm to segment the training spectral data set and divide the data set into multiple clusters with unique spectral characteristics. Each cluster is represented by a centroid, which is the average characteristic of the spectral data in the cluster. Determination of centroid: The centroid of each cluster is the representative of the spectral data in the cluster and is determined by optimizing the objective function of the clustering algorithm to ensure that the spectral characteristics of the data points in the cluster are highly similar to those of the centroid; Calculation of spectral similarity measure SOC: Calculating SOC measures the similarity between a given spectrum and the centroid. The overlap is directly measured by comparing the minimum and maximum values ​​of the spectrum at each wavelength. The closer the calculated SOC value is to zero, the higher the spectral similarity. Generation of control vector: The calculated SOC values ​​are formed into a vector, which describes the similarity information between the input spectrum and the predefined spectrum category, capturing the details and overall characteristics of the spectrum; Construction of comprehensive control vector: The control vector is concatenated with the original spectral data to generate a comprehensive control vector, which is used as the input of the generator and the discriminator to guide the metasurface design process.

5. The hypersurface inverse design method based on the anchor point network controlled generative adversarial network according to any one of claims 1 to 4, characterized in that: In step S3, the design of the generator specifically includes: A design generator is configured to receive a control vector and a random noise vector as input, wherein the control vector contains original spectral data and cluster similarity information, and the random noise vector introduces randomness into the design space; The network architecture of the generator converts the input vector into the spatial pattern of the hypersurface structure through a series of deconvolution layers, so that the generated hypersurface is similar to the real sample in structure and aligned with the target spectrum in spectral response; The generator produces a metasurface structure as a candidate structure in the design space, which satisfies the set spectral property requirements; Construct a comprehensive loss function for the generator, which consists of three parts: adversarial loss, spectral loss, and structural loss, to balance the generator's multiple optimization objectives when generating metasurface designs; Among them, the adversarial loss is used to measure the probability that the hypersurface structure generated by the generator is judged as real by the discriminator, ensuring that the generator can generate high-quality and high-fidelity designs; Among them, the spectral loss uses the spectral similarity metric SOC to evaluate the similarity between the generated metasurface spectral response and the target spectrum; Among them, the structural loss uses the structural similarity index SSIM to measure the similarity between the generated hypersurface structure and the reference structure; Different weight parameters are set to balance the impact of adversarial loss, spectral loss, and structural loss in the overall optimization process.

6. The hypersurface inverse design method based on the anchor point network controlled generative adversarial network according to any one of claims 1 to 5, characterized in that: In step S3, the design of the discriminator specifically includes: Design a discriminator to evaluate whether the generated hypersurface design conforms to the distribution of real data; The discriminator network architecture analyzes the input hypersurface structure level by level through a series of convolutional layers to extract features that distinguish real from forged structures; Construct the discriminator loss function, including adversarial loss, mismatch loss, and spectral loss, to accurately evaluate the generated design; Among them, adversarial loss is used to evaluate the discriminator's ability to distinguish between real and generated designs, which is achieved by comparing the discriminator's scores on real samples and generated samples; Among them, the mismatch loss enhances the discriminator's ability to detect inconsistencies between input conditions and generated designs, enabling the discriminator to identify designs that do not meet expectations; Among them, the spectral loss uses the spectral similarity metric SOC to evaluate the similarity between the generated spectrum and the target spectrum.

7. The hypersurface inverse design method based on the anchor point network controlled generative adversarial network according to any one of claims 1 to 6, characterized in that: In step S3, the design of the anchor network AnchorNet specifically includes: Design AnchorNet to predict the spectral response of metasurface designs; AnchorNet is built using the bottleneck ResNet architecture; The loss function of AnchorNet is constructed based on the difference between the spectrum predicted by AnchorNet and the true spectrum obtained by electromagnetic simulation, with minimization as the optimization goal.

8. The hypersurface inverse design method based on a generative adversarial network controlled by an anchor point network according to any one of claims 1 to 7, characterized in that: The network architecture of the generator includes: The input layer receives a tensor that combines the control vector and random noise; Multi-layer convolution and deconvolution operations are used to generate three-dimensional hypersurface images from high-dimensional space, with batch normalization and Leaky ReLU activation function after each layer; The output layer uses the Tanh activation function to make the pixel value of the generated image in the range of [-1, 1], which is consistent with the standardization of the input data; The network architecture of the discriminator includes: an input layer that receives the metasurface image for evaluating structural and spectral properties; Multi-layer convolutional layers are used to extract input image features, using batch normalization and Leaky ReLU activation functions; The fully connected layer maps the features extracted by the convolutional layer into a one-dimensional vector and outputs a probability value through the Sigmoid activation function, which represents the image authenticity score; The network framework of the anchor network AnchorNet includes: An input layer that receives the metasurface image for spectral response prediction; The residual bottleneck block is used for deep feature extraction, retaining important information through convolutional layers, batch normalization and ReLU activation functions, and skip connections; multiple bottleneck blocks perform convolution operations of different scales to extract and retain image features; The output layer outputs a one-dimensional spectral vector. The extracted features are mapped to the spectral response space through the convolution layer and the fully connected layer. The activation function is selected according to the requirements of the spectrum to make the output spectrum value range meet the expectations.

9. The hypersurface inverse design method based on a generative adversarial network controlled by an anchor point network according to any one of claims 1 to 8, characterized in that: In step S4, the training process of AcGAN includes: The training data includes reference metasurface, target spectral samples, and control vectors; Initialization operation, parameter initialization of the generator and discriminator, including random initialization of weights; AnchorNet pre-training, which is pre-trained to predict the spectral characteristics of metasurface designs; Training iteration: shuffling data, sampling noise samples and control vectors, and obtaining sample pairs for each training round; Discriminator update, which performs multiple update steps of the discriminator, including generating the hypersurface, evaluating the output of the reference and generated hypersurfaces, and the output of the mismatched data; Discriminator loss calculation: calculate the discriminator loss according to a specific formula and update the discriminator parameters accordingly; Generator update, generating the hypersurface and calculating the generator loss, followed by updating the generator parameters; After training is completed, the trained discriminator and generator models are output.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the method for inverse design of a hypersurface based on a generative adversarial network controlled by an anchor point network as described in any one of claims 1 to 9 is implemented.

Images