A Metasurface Inverse Design Method Based on Generative Adversarial Networks Controlled by Anchor Point Networks

Abstract

A metasurface inverse design method based on a generative adversarial network (GAN) controlled by an anchor network is proposed. The method involves collecting metasurface design parameters and spectral response data, preprocessing them, and encoding them into a format that can be handled by a deep learning model. K-means clustering is used to analyze the training data to generate structured control vectors. An AcGAN framework is constructed, integrating the spectral similarity metric SOC into the loss functions of the generator and discriminator to enhance spectral matching accuracy. Through adversarial training, the generator is trained to produce designs that conform to the target spectral characteristics, while the discriminator and the AnchorNet are optimized for rapid prediction and evaluation of the spectral response. The trained generator, combined with the control vectors and random noise, generates candidate designs. The candidate designs are then screened to select the optimal design. This invention effectively improves the spectral control accuracy and design efficiency of metasurface design, solving the challenges faced by traditional methods in handling complex spectral requirements.

Description

Technical Field

[0001] This invention relates to metasurface design technology, and in particular to a metasurface reverse design method based on a generative adversarial network controlled by an anchor network. Background Technology

[0002] In recent years, with the rapid development of photonics technology, metasurfaces have become a research hotspot in optics and photonics due to their unique ability to manipulate electromagnetic wave characteristics at the nanoscale. Through structural design, metasurfaces can precisely control the amplitude, phase, and polarization state of light at the subwavelength scale, demonstrating enormous 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. While these methods can generate structures with specific functions, they often fall short when facing complex spectral requirements, especially when designing metasurfaces with complex spectral properties, where current methods struggle to simultaneously meet the demands of spectral accuracy and computational efficiency.

[0003] Typical metal-insulator-metal (MIM) metasurfaces exhibit broad Lorentz absorption spectra due to plasmon resonances generated at the metal-dielectric interface, making them advantageous for applications in thermal radiation control and photothermal energy conversion. However, these metasurfaces suffer from overly broad spectral responses, hindering precise spectral control during design. Meanwhile, hybrid dielectric metasurfaces generate Fano resonances through subwavelength cavity resonances, exhibiting uniquely sharp spectral characteristics that make them excellent in high-precision optical sensing. However, the design process for such structures is complex and highly sensitive to parameters, easily leading to discrepancies between designed parameters and the target spectrum. While existing reverse engineering methods can alleviate these problems to some 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 fail to capture key details in spectral characteristics, such as resonance peaks and specific absorption bands, resulting in significant deficiencies in the accuracy of spectral matching of the generated metasurfaces.

[0004] Despite these challenges, existing technologies still have significant limitations in the fine-tuning of spectral properties and efficient design space exploration for metasurface design. Traditional design methods struggle to balance accuracy and efficiency when handling 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 background section above is only for understanding the background of this application, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The main objective of this invention is to provide a metasurface inverse design method based on a generative adversarial network controlled by an anchor network, thereby solving the problems of insufficient spectral modulation accuracy and low design efficiency in the prior art.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

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

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

[0010] S2. Use the K-means clustering algorithm to perform cluster analysis on the training data to generate structured control vectors related to specific spectral categories;

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

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

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

[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 the 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] The design parameters of the MIM metasurface are encoded, specifically including:

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

[0018] Among them, the plasma frequency of the metal resonator is encoded in the red channel of the RGB image, and this frequency determines the optical properties of the metal nanostructure.

[0019] The thickness of the dielectric layer is encoded in the blue channel of the RGB image, and this thickness modulates the plasma resonance frequency, affecting the absorption spectrum.

[0020] In particular, the green channel of the RGB image is not used during the encoding process and is fixed at zero;

[0021] The design parameters of hybrid dielectric metasurfaces are coded, specifically including:

[0022] The design parameters of the hybrid dielectric structure are encoded using the RGB image format, with each pixel representing a unit 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, which serves 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, and this thickness affects the resonant behavior in the hybrid dielectric structure.

[0025] In particular, the red channel of the RGB image is not used during the encoding process and is fixed at zero.

[0026] Furthermore, 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 quantities to a uniform range.

[0027] Furthermore, step S2 specifically includes:

[0028] Spectral data clustering: The training spectral dataset is segmented using a clustering algorithm, dividing the dataset into multiple clusters with unique spectral characteristics. Each cluster is represented by a centroid, which is the average feature of the spectral data in that cluster.

[0029] Centroid determination: The centroid of each cluster is determined by optimizing the objective function of the clustering algorithm, which serves as a representative of the spectral data within that cluster, to ensure that the spectral characteristics of the data points within the cluster are highly similar to those of the centroid.

[0030] Calculation of Spectral Similarity Measure SOC: The spectral similarity measure SOC is calculated 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 SOC value is to zero, the higher the spectral similarity.

[0031] Control vector generation: 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 features of the spectrum.

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

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

[0034] The design generator takes a control vector and a random noise vector as input. The control vector contains the original spectral data and cluster similarity information, while the random noise vector introduces randomness into the design space.

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

[0036] The generator produces metasurface structures as candidate structures in the design space, which meet the set spectral characteristic requirements;

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

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

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

[0040] Among them, the structural loss is measured by the structural similarity index (SSIM) to measure the similarity between the generated metasurface structure and the reference structure;

[0041] Different weighting parameters are set to balance the effects 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 metasurface design conforms to the distribution of real data;

[0044] The discriminator network architecture analyzes the input metasurface structure step by step through a series of convolutional layers to extract features that distinguish between real and fake structures;

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

[0046] Among them, the 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, spectral loss uses the spectral similarity measure (SOC) to evaluate the similarity between the generated spectrum and the target spectrum.

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

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

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

[0052] Construct a loss function for AnchorNet that minimizes the difference between the spectrum predicted by AnchorNet and the true spectrum obtained through electromagnetic simulation.

[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 reduce the dimensionality of high-dimensional space to generate 3D metasurface images, with each layer followed by batch normalization and Leaky ReLU activation function;

[0056] The output layer uses the Tanh activation function to ensure that the pixel values ​​of the generated image are 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] The input layer receives images of the metasurface for evaluating its structural and spectral properties.

[0059] Multiple convolutional layers are used to extract features from the input image, employing batch normalization and the Leaky ReLU activation function;

[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, representing the image authenticity score.

[0061] The AnchorNet network framework includes:

[0062] The input layer receives images of the metasurface and is used for spectral response prediction.

[0063] Residual bottleneck blocks are used for deep feature extraction. They preserve important information through convolutional layers, batch normalization, ReLU activation functions, and skip connections. Multiple bottleneck blocks perform convolutional operations of different scales to extract and preserve image features.

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

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

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

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

[0068] AnchorNet pre-training is used to predict the spectral properties of metasurface designs.

[0069] The training iteration involves shuffling the data, sampling noise samples and control vectors, and obtaining sample pairs for each training round.

[0070] Discriminator update: Perform multiple update steps for the discriminator, including generating a metasurface, evaluating the output of the reference and generated metasurfaces, and outputting mismatched data;

[0071] Discriminator loss calculation: The discriminator loss is calculated according to a specific formula, and the discriminator parameters are updated accordingly.

[0072] The generator is updated by generating a metasurface and calculating the generator's loss, followed by updating the generator parameters.

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

[0074] In a second aspect of the invention, a computer program product includes a computer program that, when executed by a processor, implements the metasurface reverse design method based on anchor network control and generative adversarial networks.

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

[0076] This invention proposes a metasurface inverse design method based on Generative Adversarial Networks (AcGAN), providing an innovative design framework. By introducing an innovative spectral similarity metric (SOC) and structured control vectors, it significantly improves the spectral matching accuracy and design efficiency of metasurface design. This metasurface inverse design method can accurately capture spectral features, improve design efficiency, and is applicable to various metasurface types.

[0077] First, this invention innovatively proposes introducing a spectral similarity measure (SOC) as part of the loss function evaluation, building upon existing mean squared error (MSE) and mean absolute error (MAE), to accurately measure the degree of overlap between the generated and target spectra. By calculating the ratio of the minimum and maximum values ​​between the spectra, SOC can better capture key spectral features such as resonance peaks and specific absorption bands, significantly improving the accuracy of metasurface spectral manipulation. This innovation directly addresses the limitation of traditional loss functions in capturing spectral details, ensuring more accurate spectral matching of the generated metasurface.

[0078] Secondly, this invention introduces structured control vectors into the design framework. The training spectral data is segmented and clustered using the K-means clustering algorithm to generate control vectors associated with specific spectral categories. These control vectors are then combined with the original spectral data and input into the generator and discriminator. This two-layer data processing method not only preserves the detailed information of the spectrum but also enhances the model's exploration capability 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 handling complex spectral characteristics, enabling the generated metasurface design to meet stringent spectral requirements in various application scenarios.

[0079] Furthermore, this invention utilizes the AnchorNet module within the AcGAN framework to rapidly predict the spectral response of metasurface designs, reducing computation time during design iterations and improving overall design efficiency. The AnchorNet not only provides accurate spectral response predictions during adversarial generation but also prevents model overfitting through an early stopping mechanism, ensuring the broad applicability and high efficiency of the generated metasurface designs. These innovations comprehensively enhance the design capabilities of metasurfaces, providing stronger technical support for advanced applications in photonics.

[0080] The innovative design method of this invention overcomes the shortcomings of existing technologies in terms of fine-tuning of spectral characteristics and efficient design exploration, 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. Attached Figure Description

[0082] Figure 1 This is a diagram of the AcGAN model framework 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 AcGAN framework in this embodiment of the invention: generator, discriminator, and AnchorNet. Detailed Implementation

[0084] The embodiments of the present invention will be described in detail below. It should be emphasized that the following description is merely exemplary and not intended to limit the scope and application of the present invention.

[0085] This invention provides a metasurface inverse design method based on a generative adversarial network controlled by an anchor network, which mainly includes the following steps:

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

[0087] S2. Use the K-means clustering algorithm to perform cluster analysis on the training data to generate structured control vectors related to specific spectral categories;

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

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

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

[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] This invention provides an innovative metasurface inverse design method based on an anchor network-controlled generative adversarial network (GAN). This method significantly improves the accuracy and efficiency of metasurface design through several key innovations. First, this invention introduces the spectral similarity metric SOC as a means to evaluate the similarity between the generated and target spectra, integrating it into the loss functions of both the generator and discriminator. This allows the loss function to accurately measure the similarity between the generated and target spectra, effectively capturing key details such as resonance peaks and specific absorption bands, thus overcoming the shortcomings of traditional loss functions in fine-grained spectral control. Second, this invention introduces a structured control vector and utilizes the K-means clustering algorithm to effectively segment and cluster the training data, enhancing the model's ability to explore the design space while avoiding error amplification under complex spectral requirements. Furthermore, this invention constructs an AnchorNet module within the AcGAN framework, enabling rapid prediction of the spectral response of the metasurface design, greatly reducing the computation time of design iterations and improving the efficiency of the design process. This 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 following describes specific embodiments of the present invention.

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

[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 deep learning models 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 spectral responses.

[0096] MIM metasurface coding:

[0097] MIM structures exhibit a broad Lorentz absorption spectrum due to plasmon resonance at the metal-dielectric interface. These structures are commonly used in fields such as thermal radiation control and photothermal energy conversion. 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 nm × 50 nm unit cell. The specific mapping is as follows:

[0098] • Red channel (R): Plasma frequency ω encoding the metal resonator p Plasma frequency is a key parameter that determines the optical properties of metallic nanostructures, affecting the resonance wavelength and absorption characteristics.

[0099] • Blue channel (B): Encodes the thickness d of the dielectric layer, in nanometers. The thickness of the dielectric layer modulates the plasma resonance frequency, directly affecting the absorption spectrum.

[0100] • Green Channel (G): This channel is unused and is fixed at 0 to ensure the simplicity and consistency of the encoding.

[0101] Hybrid dielectric metasurface coding:

[0102] Hybrid dielectric metasurfaces generate Fano resonances through subwavelength cavity resonances, exhibiting sharp spectral characteristics that demonstrate excellent performance in high-precision optical sensing. To encode the design parameters of this structure into a form processable by deep learning models, a 3×64×64 RGB image format was used, where each pixel represents a 120 nm × 120 nm unit cell. The specific mapping is as follows:

[0103] • Green Channel (G): Encodes the real refractive index n of the dielectric resonator. The refractive index is a core parameter affecting the optical performance of the dielectric and directly influences the characteristics of the Fano resonance.

[0104] • Blue channel (B): Encodes the thickness d of the dielectric layer in nanometers. This parameter is the same as that in the MIM structure and is crucial for the resonant behavior in the hybrid dielectric structure.

[0105] • Red channel (R): This channel is unused and is fixed at 0.

[0106] Spectral discretization and normalization:

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

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

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

[0110] This encoding and data preprocessing method provides a unified input format for deep learning models, enabling the optimization and analysis of different types of metasurface designs within the same framework. The RGB image encoding method allows the model to utilize convolution operations in convolutional neural networks to capture the spatial and spectral correlations of design parameters. Simultaneously, spectral discretization and normalization ensure the consistency and stability of the model's input data, thereby improving the model's prediction accuracy and generalization ability. In practical applications, this method is not only applicable to the design of MIM and hybrid dielectric metasurfaces but can also be extended to the design of other types of nanostructures as needed. By adjusting the pixel scale of the encoding and introducing more physical parameters (such as dielectric constant and anisotropy properties), 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 possesses sufficient accuracy and efficiency when handling complex spectral responses and design tasks, meeting diverse application needs.

[0111] 2. Clustering of Spectral Data and Generation of Control Vectors

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

[0113] Clustering of spectral data:

[0114] First, to address the design problem of metasurfaces, a training spectral dataset S = {s1, s2, ..., s} was obtained. n}, where each s i This represents a data point with a unique spectral response. To efficiently process this spectral data and extract representative spectral features, the K-means clustering algorithm is used to segment and cluster the spectral data.

[0115] The goal of the K-means clustering algorithm is to divide a spectral dataset S into k distinct clusters C = {C1, C2, ..., Ck}. k The clustering results are optimized by minimizing the following objective function:

[0116]

[0117] Among them, c i Represents the i-th cluster C iThe centroid (i.e., the cluster center) represents the average characteristic of the spectral data in that cluster.

[0118] Each cluster C is clustered using K-means. i The spectral data points in the data are all related to the centroid c of the cluster. i The clusters are designed to be as close as possible in the spectral feature space to ensure that the spectral data within each cluster have highly similar spectral characteristics. After clustering, k centroids are obtained, each of which can be considered as a representative of a certain class of spectral features.

[0119] Calculation of SOC (Spectral Similarity Measure):

[0120] Obtaining the centroid c i Subsequently, a novel spectral similarity measure (SOC) was proposed to measure the similarity between a given spectrum and these centroids. The SOC is calculated as follows:

[0121]

[0122] Where s represents a specific spectral response, c i This represents the i-th centroid. The closer the SOC value is to zero, the closer s is to c. i The higher the spectral similarity, the better.

[0123] SOC (Signal Occlusion) directly measures the degree of 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 portion of the entire spectrum but are crucial.

[0124] Generation of control vectors:

[0125] Once the spectrum s and each centroid c are calculated... i The SOC values ​​between them are represented as a k-dimensional vector:

[0126] v=[SOC(s,c1),SOC(s,c2),…,SOC(s,c k (3)

[0127] This vector v quantitatively describes the similarity information between the input spectrum s and k predefined spectral categories (i.e., centroids). This similarity vector not only contains detailed information about the spectrum but also captures the overall characteristics of the spectrum through clustering.

[0128] Next, the vector v is concatenated with the original spectral data s to generate a comprehensive control vector u:

[0129] u = [s; v] (4)

[0130] The control vector u combines the original spectral information and the clustered similarity information, serving as input to the generator and discriminator to guide the subsequent metasurface design process.

[0131] Application of control vectors in generators and discriminators:

[0132] In the generator, the control vector u is used as input to generate metasurface structures that meet specific spectral requirements. By combining the raw spectral data with clustering information, the generator can better capture important features in the spectrum and generate design schemes that better meet the target requirements.

[0133] In the discriminator, the control vector u is used to evaluate whether the generated metasurface design meets the expected spectral response. Because the control vector contains clustering information, the discriminator can more accurately identify subtle differences between the generated design and the target spectrum, thereby improving design quality.

[0134] Overall significance:

[0135] By introducing clustering of spectral data and control vector generation, this invention significantly improves the spectral matching accuracy and design efficiency in the metasurface design process. SOC, as a novel spectral similarity metric, effectively addresses the problem that traditional loss functions cannot capture spectral details. Furthermore, through the generation and application of control vectors, the entire design process can more effectively explore the design space and ensure that the generated metasurface structures exhibit high accuracy and consistency in spectral response. This method is not only applicable to current metasurface design needs but can also be extended to the design of other spectrally complex nanostructures by adjusting the clustering algorithm and control vector generation method, supporting a wider range of photonics applications.

[0136] 3. AcGAN-based metasurface design and generation

[0137] In metasurface design, to achieve precise matching of complex spectral responses, a generative method called Anchor-controlled Generative Adversarial Network (AcGAN) is proposed. The AcGAN framework is an innovative network structure proposed in this invention, integrating four core components: controller, generator, discriminator, and anchor network (AnchorNet). Through adversarial training, it optimizes design parameters to generate metasurface structures that meet specific spectral requirements.

[0138] AcGAN Framework Overview:

[0139] The core of the AcGAN framework lies in continuously optimizing the generated metasurface designs through adversarial training between the generator and discriminator. The generator is responsible for generating metasurface structures with realistic spectral properties 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. The AcGAN framework, as shown below... Figure 1 As shown.

[0140] Generator design and input:

[0141] The generator receives a control vector *u* and a random noise vector *z* as input. The control vector *u* contains the original spectral data and cluster similarity information, while the noise vector *z* introduces randomness into the design space, allowing the generator to explore diverse design schemes. The generator transforms these inputs into a spatial pattern of the metasurface structure through a series of deconvolutional layers. The generator's main task is to ensure that the generated metasurface is not only structurally similar to the real sample but also spectrally aligned with the target spectrum.

[0142] Generator output and loss function:

[0143] The generator output is a metasurface structure M g This represents the candidate structures generated in the design space. To quantify the generator's performance, a comprehensive loss function L is proposed. G It consists of the following three parts:

[0144]

[0145] • Combating losses This measures whether the generator successfully fooled the discriminator. The specific formula is:

[0146]

[0147] Where D(G(z,u),u) is the discrimination score given by the discriminator, representing the probability that the generated design is judged as real.

[0148] ·Spectral loss L spectral This paper proposes using a spectral similarity measure (SOC) to evaluate the spectral response of the generated metasurface. g With the target spectrum s t The similarity between them. The specific formula is:

[0149] L spectral =SOC(s g ,s t (7)

[0150] Structural loss L structural: Proposes using the Structural Similarity Index (SSIM) to measure the M of the generated metasurface structure g With reference structure M r The similarity between them. The specific formula is:

[0151] L structural =SSIM(M g M r (8)

[0152] 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, respectively. The generator minimizes the comprehensive loss L... G We continuously improve its output so that the generated metasurface design not only meets spectral requirements but also maintains structural rationality.

[0153] Design and function of the discriminator:

[0154] The discriminator is a key component of the AcGAN framework. It is responsible for evaluating the generated metasurface design and determining whether it conforms to the distribution of real data. The discriminator analyzes the input metasurface structure step by step through a series of convolutional layers, thereby extracting features that can distinguish between real and fake structures.

[0155] The loss function L of the discriminator D It includes multiple parts to ensure that it can accurately evaluate the generated design:

[0156]

[0157] • Combating losses Evaluate the discriminator's ability to distinguish between real and generative designs.

[0158] The specific formula is as follows:

[0159]

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

[0161] Mismatch loss L mismatch This paper proposes an enhanced discriminator's ability to detect inconsistencies between input conditions and the generative design. 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 L spectral Similar to the generator, the discriminator evaluates the generated spectra using SOC. g With the target spectrum s t The similarity between them.

[0165] The introduction of AnchorNet:

[0166] AnchorNet is a module proposed and designed within the AcGAN framework for fast spectral response prediction. It reduces computational complexity and optimizes spectral prediction accuracy through a bottleneck ResNet architecture. The primary goal of AnchorNet is to minimize the predicted spectral response. The difference between the spectrum and the true spectrum s is optimized using the following loss function:

[0167]

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

[0169] The training process of AcGAN:

[0170] To enable the generator and discriminator to generate and evaluate high-quality metasurface designs, the following training procedure is proposed:

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

[0172] Algorithm 1: AcGAN Training Process for Metasurface Design Input: Training pairs in For reference metasurface, For the target spectral sample, u i This is the control vector.

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

[0174] 1. Initialize the parameters θ of generator G G The parameters θ of the discriminator D D Randomly initialize the weights.

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

[0176] 3. For each training epoch from epoch = 1 to E, perform the following operations:

[0177] 4. Shuffle the training data;

[0178] 5. From the noise distribution p Z m noise samples are obtained by sampling from (z) {z (1) ,…,z (m)}; From the control vector distribution p U m control vectors {u} are sampled from (u) (1) ,…,u (m)}; From the data distribution p data (M r ,s t m sample pairs were obtained by sampling from )

[0179] 6. For each batch {z,u,x} of size B, perform the following operation: Update the discriminator.

[0180] 7. Perform the discriminator update steps k times:

[0181] 8.M g ←G(z,u) / / Generates a metasurface with condition u;

[0182] 9.D real ←D(x,u) / / Output of the discriminator to the reference metasurface;

[0183] 10.D fake ←D(M g ,u) / / The discriminator's output for generating the metasurface;

[0184] 11.D mismatch ←D(M g ,u′) / / The output of the discriminator for mismatched data;

[0185] 12. Calculate the discriminator loss L D :

[0186]

[0187] 13. Update the discriminator parameters θ D :

[0188]

[0189] 14. End the discriminator update step.

[0190] Update Generator

[0191] 15.M g ←G(z,u) / / Generates a metasurface with condition u;

[0192] 16.D fake ←D(M g ,u) / / The discriminator's output for generating the metasurface;

[0193] 17. Calculate the generator loss L G :

[0194]

[0195] 18. Update generator parameter θ G :

[0196]

[0197] Output: The trained discriminator D and generator G models.

[0198] Through this iterative training, the AcGAN framework can 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.

[0199] Overall significance:

[0200] By proposing and applying the AcGAN framework, this invention achieves efficient generation and optimization of metasurface designs. The generator can explore a wide design space, the discriminator provides rigorous design quality control, and AnchorNet ensures the accuracy of spectral predictions. This integrated generation method enables the design of metasurfaces with complex spectral requirements, greatly improving design efficiency and accuracy.

[0201] 4. Network Architecture

[0202] In the AcGAN framework, the generator, discriminator, and AnchorNet are the three core components, each undertaking a specific task and achieving these tasks through a carefully designed network architecture. The following is a detailed description of the network architecture of these three components: [Detailed description follows] Figure 2 As shown. Generator.

[0203] The generator's main task is to receive a control vector u and a random noise vector z, and generate a realistic metasurface structure M through a series of deconvolution layers. g The network architecture of the generator is shown in Figure (a), and it contains several key components:

[0204] • Input layer: The generator's input is a tensor of size 2000×1×1, containing a combination of control vector u and random noise z.

[0205] • Convolutional and deconvolutional layers: The generator uses multiple layers of convolutional and deconvolutional operations to gradually reduce the dimensionality from high-dimensional space and generate the final 3D metasurface image. Each layer is accompanied by batch normalization (BatchNorm) and LeakyReLU activation functions to stabilize the training process and improve the quality of the generated images.

[0206] • Output layer: The generator output is a tensor of size 3×64×64, representing the generated metasurface structure M. g The output layer uses the Tanh activation function to ensure that the pixel values ​​of the generated image are in the range of [-1, 1], which is consistent with the normalization of the input data.

[0207] Discriminator

[0208] The discriminator's task is to receive the generated metasurface structure M g and reference metasurface structure M r They are evaluated and their authenticity is determined. The architecture of the discriminator is shown in Figure (b), and it has the following characteristics:

[0209] • Input layer: The discriminator takes a 3×64×64 metasurface image (generated or referenced) as input for evaluating its structural and spectral properties.

[0210] • Convolutional layer: The discriminator extracts features from the input image step by step through multiple convolutions. The convolution operation is accompanied by batch normalization and LeakyReLU activation function to ensure that the discriminator can extract important spatial features.

[0211] • 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.

[0212] AnchorNet

[0213] 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, as shown in Figure (c), which includes the following key components:

[0214] • Input layer: The input to AnchorNet is a 3×64×64 metasurface image (generated or referenced), which is the same as the input to the discriminator.

[0215] • Residual Bottleneck Blocks: AnchorNet uses residual bottleneck blocks to process input data. These bottleneck blocks extract deep features using convolutional layers, batch normalization, and ReLU activation functions, and preserve important information in the image through skip connections (ShortCut). The network contains multiple bottleneck blocks, each performing convolutional operations of different scales to extract features.

[0216] • 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 convolutional and fully connected layers. The activation function of the output layer is selected based on the specific spectral requirements (such as Tanh or Sigmoid) to ensure that the numerical range of the output spectrum meets expectations.

[0217] These three network components work collaboratively to achieve an efficient and accurate metasurface design and generation process within the AcGAN framework. The generator continuously improves its ability to generate metasurface structures through adversarial training against the discriminator and anchor network. The discriminator provides feedback to improve the generator's performance by evaluating the realism and spectral consistency of the generated structures. The anchor network provides additional optimization guidance to the generator by rapidly and accurately predicting the spectral response.

[0218] Through these carefully designed network architectures, AcGAN can effectively handle complex metasurface design tasks, ensuring that the generated designs have high-quality spectral properties and structural rationality in practical applications.

[0219] 5. Design optimization and iterative improvement

[0220] In the AcGAN framework, after multiple rounds of iterative training, the generator outputs a series of candidate metasurface designs. These designs require further design exploration and optimization to ensure their performance in practical applications. Part 5 will describe in detail how to screen, test, and ultimately optimize these candidate designs.

[0221] Design exploration and candidate solution generation

[0222] After multiple rounds of training, the generator not only outputs a single design solution but also generates a set of candidate designs with diverse characteristics. These designs reflect the various possibilities captured by the generator while exploring the design space. Specifically, these include:

[0223] • Design of different structural parameters: Since there are many possible combinations of control vector u and random noise vector z, the generator can output metasurface designs with various combinations of structural parameters to adapt to different spectral requirements.

[0224] • Diverse spectral responses: The designs generated by the generator may exhibit different characteristics in their spectral responses, such as broadband absorption, narrowband selective absorption, or high reflectivity in specific wavelength bands. These spectral characteristics enable candidate designs to cover a wider range of application scenarios.

[0225] Screening of candidate designs

[0226] Selecting the optimal solution from the candidate designs output by the generator is a crucial step in achieving efficient application. This process typically includes the following steps:

[0227] Spectral performance screening: The spectral response of each candidate design is accurately calculated using electromagnetic simulation tools (such as LumericalFDTD) and compared with the target spectrum. The designs that perform best on key spectral parameters are selected.

[0228] Structural reliability assessment: This involves conducting stability tests on the physical structure of candidate designs to evaluate their feasibility in actual manufacturing processes and structural integrity under various environmental conditions. This includes simulating factors that may affect design performance, such as thermal effects, mechanical stress, and material aging.

[0229] Real-world application testing: The selected designs need to be tested in a real-world application environment to verify whether their performance meets expectations. Real-world application testing includes:

[0230] • Physical fabrication and prototype testing: The candidate design is manufactured into a physical sample and tested under laboratory conditions. The actual performance of the design is verified by measuring actual parameters such as spectral response, reflectance, and transmittance.

[0231] • Environmental adaptability testing: The prototype design is tested under different environmental conditions (such as high temperature, humidity, mechanical stress) to ensure its long-term stability and reliability in different application scenarios.

[0232] The final choice for design optimization

[0233] Following practical application testing, the final design selection is based not only on precise matching of the spectral response but also on comprehensive considerations such as manufacturing costs, material selection, and structural complexity. The final selected design will undergo further optimization to meet specific application requirements. The optimization process may include:

[0234] • Material optimization: Based on actual manufacturing needs, optimize material selection to reduce costs while maintaining high performance in the design.

[0235] • Process improvement: Based on the limitations of the manufacturing process, the design is appropriately simplified or adjusted to improve production efficiency and yield.

[0236] Feedback and Continuous Improvement

[0237] The final selected design may still require fine-tuning and improvement during practical application. This process feeds the generated design back into the AcGAN training framework for secondary optimization, enabling the generator to output design solutions that better meet application requirements in subsequent design tasks.

[0238] In summary, 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 excellently in practical applications. This process provides strong support for future complex photonics and optics applications.

[0239] In summary, this invention proposes a metasurface inverse design method based on anchor network-controlled generative adversarial networks (AcGANs). This method improves the spectral matching accuracy and design efficiency of metasurface design through the following innovations:

[0240] A spectral similarity measure (SOC) is introduced 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 spectra at each wavelength, key features in the spectrum, such as resonance peaks and specific absorption bands, are effectively captured, thereby improving the accuracy of spectral modulation.

[0241] Structured control vectors were adopted, and the training spectral data was analyzed using the K-means clustering algorithm to generate control vectors associated with specific spectral categories. These vectors were combined with the original spectral data and input into the generator and discriminator, which enhanced the model's ability to explore the design space and preserved the detailed information of the spectrum.

[0242] By utilizing the AnchorNet module in the AcGAN framework, we can quickly predict the spectral response of metasurface designs, reduce the computation time of design iterations, and prevent model overfitting through an early stopping mechanism, thereby improving the wide applicability and efficiency of the designs.

[0243] By comprehensively utilizing the above-mentioned innovations, the metasurface reverse design method of this invention not only improves the spectral matching accuracy of the design, but also significantly increases the design efficiency. It is applicable to a variety of metasurface types and provides a solid technical foundation for advanced applications in photonics and the development of next-generation optical devices and sensing technologies.

[0244] This invention also provides a storage medium for storing a computer program, which, when executed, performs at least the methods described above.

[0245] This invention also provides a control device, including a processor and a storage medium for storing a computer program; wherein the processor executes the computer program by performing at least the method described above.

[0246] This invention also provides a processor that executes a computer program, at least performing the methods described above.

[0247] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. 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 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 storage device or a magnetic tape storage device. The storage media described in the embodiments of this invention are intended to include, but are not limited to, these and any other suitable types of memory.

[0248] In the several embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0249] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0250] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0251] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0252] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0253] The methods disclosed in the several method embodiments provided by this invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0254] The features disclosed in the several product embodiments provided by this invention can be arbitrarily combined without conflict to obtain new product embodiments.

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

[0256] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or application, should be considered within the scope of protection of the present invention.

Claims

1. A metasurface inverse design method based on generative adversarial networks controlled by anchor networks, characterized in that, Includes the following steps: S1. Determine the spectral properties required for metasurface design and determine the target spectrum; Collect and preprocess training data, including collecting metasurface design parameters and corresponding spectral response data, and preprocessing them to adapt them to the deep learning model, encoding the metasurface design parameters into a format that the deep learning model can process; S2. Use the K-means clustering algorithm to perform cluster analysis on the training data to generate structured control vectors related to specific spectral categories; Step S2 specifically includes: Spectral data clustering: Clustering algorithms are used to segment the training spectral dataset, dividing the dataset into multiple clusters with unique spectral characteristics, each represented by a centroid; Centroid determination: The centroid of each cluster is determined by optimizing the objective function of the clustering algorithm, which serves as a representative of the spectral data within that cluster, to ensure that the spectral characteristics of the data points within the cluster are highly similar to those of the centroid. Calculation of Spectral Similarity Measure SOC: The SOC measures the similarity between a given spectrum and its centroid. The closer the calculated SOC value is to zero, the higher the spectral similarity. Control vector generation: 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 features of the spectrum; Construction of the integrated control vector: The integrated control vector is generated from the control vector and the original spectral data. This integrated control vector is used as the input to the generator and discriminator to guide the metasurface design process. S3. Construct the AcGAN framework, which includes a generator, a discriminator, and an anchor network AnchorNet. In this framework, the spectral similarity metric SOC is introduced as a means to evaluate the similarity between the generated spectrum and the target spectrum and is integrated into the loss function of the generator and the discriminator. S4. Use adversarial training to train the generator to generate metasurface designs that conform to the target spectral characteristics, while training the discriminator and AnchorNet to quickly predict and evaluate the spectral response; S5. Using the trained generator, generate a series of candidate metasurface designs based on the control vector and random noise; S6. Use evaluation tools to screen candidate designs and select the optimal design.

2. The metasurface inverse design method based on generative adversarial networks controlled by anchor networks as described 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 the data is preprocessed. The design parameters of the MIM metasurface are encoded, specifically including: The design parameters of the MIM structure are encoded using the RGB image format, with each pixel representing a unit of a predetermined size; Among them, the plasma frequency of the metal resonator is encoded in the red channel of the RGB image, and this frequency determines the optical properties of the metal nanostructure. The thickness of the dielectric layer is encoded in the blue channel of the RGB image, and this thickness modulates the plasma resonance frequency, affecting the absorption spectrum. In particular, the green channel of the RGB image is not used during the encoding process and is fixed at zero; The design parameters of hybrid dielectric metasurfaces are coded, specifically including: The design parameters of the hybrid dielectric structure are encoded using the RGB image format, with each pixel representing a unit of a predetermined size. Among them, the real refractive index of the dielectric resonator is encoded in the green channel of the RGB image, which serves 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, and this thickness affects the resonant behavior in the hybrid dielectric structure. In particular, the red channel of the RGB image is not used during the encoding process and is fixed at zero.

3. The metasurface inverse design method based on generative adversarial networks controlled by anchor networks as described in 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, with 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 quantities to a uniform range.

4. The metasurface reverse design method based on anchor network control of generative adversarial networks as described in any one of claims 1 to 2, characterized in that, Step S2 also includes: The centroid is the average feature of the spectral data in this cluster; The calculation of SOC includes directly measuring the degree of overlap by comparing the minimum and maximum values ​​of the spectrum at each wavelength; The control vector is concatenated with the original spectral data to generate the integrated control vector.

5. The metasurface inverse design method based on generative adversarial networks controlled by anchor networks as described in any one of claims 1 to 2, characterized in that, In step S3, the generator design specifically includes: The design generator takes a control vector and a random noise vector as input. The control vector contains the original spectral data and cluster similarity information, while the random noise vector introduces randomness into the design space. The generator's network architecture transforms the input vector into a spatial pattern of the metasurface structure through a series of deconvolution layers, so that the generated metasurface is structurally similar to the real sample and aligned with the target spectrum in terms of spectral response; The generator produces metasurface structures as candidate structures in the design space, which meet the set spectral characteristic requirements; A comprehensive loss function for the generator is constructed, which consists of three parts: adversarial loss, spectral loss, and structural loss, to balance the multiple optimization objectives of the generator when generating metasurface designs; Among them, the adversarial loss is used to measure the probability that the metasurface structure generated by the generator is judged as real by the discriminator, ensuring that the generator can generate high-quality, high-fidelity designs. Among them, the spectral loss is evaluated using the spectral similarity metric SOC to assess the similarity between the generated metasurface spectral response and the target spectrum; Among them, the structural loss is measured by the structural similarity index SSIM to measure the similarity between the generated metasurface structure and the reference structure; Different weighting parameters are set to balance the effects of adversarial loss, spectral loss, and structural loss in the overall optimization process.

6. The metasurface inverse design method based on generative adversarial networks controlled by anchor networks as described in any one of claims 1 to 2, characterized in that, In step S3, the design of the discriminator specifically includes: Design a discriminator to evaluate whether the generated metasurface design conforms to the distribution of real data; The discriminator network architecture analyzes the input metasurface structure step by step through a series of convolutional layers to extract features that distinguish between real and fake structures; Construct a loss function for the discriminator, including adversarial loss, mismatch loss, and spectral loss, to accurately evaluate the generated design; Among them, the 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, spectral loss uses the spectral similarity metric SOC to evaluate the similarity between the generated spectrum and the target spectrum.

7. The metasurface reverse design method based on generative adversarial networks controlled by anchor networks as described in any one of claims 1 to 2, characterized in that, In step S3, the design of the AnchorNet specifically includes: Design AnchorNet to predict the spectral response of metasurface designs; AnchorNet is built using the bottleneck ResNet architecture; Construct a loss function for AnchorNet that minimizes the difference between the spectrum predicted by AnchorNet and the true spectrum obtained through electromagnetic simulation.

8. The metasurface inverse design method based on generative adversarial networks controlled by anchor networks as described in any one of claims 1 to 2, 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 reduce the dimensionality of high-dimensional space to generate 3D metasurface images, with each layer followed by batch normalization and Leaky ReLU activation function; The output layer uses the Tanh activation function to ensure that the pixel values ​​of the generated image are in the range of [-1, 1], which is consistent with the standardization of the input data. The network architecture of the discriminator includes: The input layer receives images of the metasurface for evaluating its structural and spectral properties. Multiple convolutional layers are used to extract features from the input image, employing batch normalization and the Leaky ReLU activation function; 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, representing the image authenticity score. The AnchorNet network framework includes: The input layer receives images of the metasurface and is used for spectral response prediction. Residual bottleneck blocks are used for deep feature extraction. They preserve important information through convolutional layers, batch normalization, ReLU activation functions, and skip connections. Multiple bottleneck blocks perform convolutional operations of different scales to extract and preserve image features. The output layer outputs a one-dimensional spectral vector. The extracted features are mapped to the spectral response space through convolutional layers and fully connected layers. The activation function is selected according to the spectral requirements to make the output spectral value range meet the expectations.

9. The metasurface inverse design method based on generative adversarial networks controlled by anchor networks as described in any one of claims 1 to 2, characterized in that, In step S4, the training process of AcGAN includes: The training data includes a reference metasurface, target spectral samples, and control vectors; Initialization operations include parameter initialization for the generator and discriminator, including random initialization of weights; AnchorNet pre-training is used to predict the spectral properties of metasurface designs. The training iteration involves shuffling the data, sampling noise samples and control vectors, and obtaining sample pairs for each training round. Discriminator update: Perform multiple update steps for the discriminator, including generating a metasurface, evaluating the output of the reference and generated metasurfaces, and outputting mismatched data; Discriminator loss calculation: The discriminator loss is calculated according to a specific formula, and the discriminator parameters are updated accordingly. The generator is updated by generating a metasurface and calculating the generator's loss, followed by updating the generator parameters. After training is complete, 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 the processor, it implements the metasurface reverse design method based on anchor network control of generative adversarial networks as described in any one of claims 1-9.

Images