Geometric damage metasurface global inversion design method based on super learning

Abstract

The invention relates to a geometric damage metasurface global inversion design method based on super learning, and belongs to the technical field of intelligent metasurface and artificial electromagnetic structure design. The method is realized through four steps of input data preparation, super-network-based generation stage, optimization stage and output verification. In the generation stage, the improved CVAE is used as a generator, a super network dynamically generates decoder parameters, and adversarial learning and MMD constraint are introduced; in the optimization stage, a full-connection neural network optimizer is adopted, and candidate patterns are corrected through greedy iteration. The method solves the problems that a traditional inversion design is poor in geometric damage robustness, fuzzy in output, insufficient in real-time performance and the like, has high robustness, high definition, high real-time performance, high precision and high adaptability, and can be widely applied to the fields of wireless communication, radar stealth and the like.

Description

Technical Field

[0001] This invention relates to a global inversion design method for metasurfaces with geometric damage (such as structural breakage, missing elements, corrosion or occlusion) based on super learning. It is particularly suitable for applications such as wireless communication, radar stealth, intelligent sensing and optical computing where the real-time performance and accuracy of electromagnetic response are highly demanding. It belongs to the field of intelligent metasurface and artificial electromagnetic design technology. Background Technology

[0002] Smart metasurfaces, as a novel type of artificial electromagnetic material capable of flexibly controlling the amplitude, phase, and polarization of electromagnetic waves, possess core value in achieving desired electromagnetic responses through specific structural designs, and have become a key support for many high-tech fields. The design of metasurfaces typically relies on inverse design methods, that is, deriving the corresponding metasurface structural parameters from the desired electromagnetic field distribution.

[0003] Existing inversion design techniques mainly rely on traditional deep learning models, such as deep neural networks (DNNs), convolutional neural networks (CNNs), and conditional variational autoencoders (CVAEs), but they have the following significant technical drawbacks: Non-unique mapping leads to fuzzy output: When traditional models face the non-unique mapping relationship between electromagnetic field distribution and metasurface structure, they tend to output averaged and fuzzy structural patterns, which cannot meet the requirements of high-precision applications.

[0004] Poor adaptability to geometric damage: In practical applications, metasurfaces inevitably suffer geometric damage, including structural breakage, missing elements, corrosion, or shading. However, traditional models assume structural integrity and do not consider geometric disturbances, leading to a significant deviation between the electromagnetic response after damage and the design expectations, resulting in direct equipment failure. Insufficient real-time performance and economy: After structural damage occurs, traditional methods require retraining the model for the specific damage scenario, which not only incurs huge computational costs but also fails to meet the requirements of real-time electromagnetic control.

[0005] Low accuracy and correlation: Traditional models have high far-field prediction errors and weak correlation with the target distribution, making it difficult to achieve precise electromagnetic control.

[0006] Therefore, developing a metasurface inversion design method that can maintain high accuracy, high robustness and real-time response under geometric damage conditions has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0007] To overcome the shortcomings of existing metasurface inversion design methods, this invention provides a global inversion design method for geometrically damaged metasurfaces based on superlearning. It combines a generative-optimization (GO) framework with a Hepernetwork dynamic parameter adjustment mechanism, which can quickly generate high-quality metasurface structures under structural damage and achieve optimization correction within milliseconds.

[0008] A global inversion design method for geometrically damaged metasurfaces based on super-learning includes the following steps: Step 1: Input Data Preparation: Obtain a dataset containing the metasurface geometry matrix, the target far-field distribution matrix, and structural damage information to form an input sample pair (F,G) and the real structural pattern P; the input data may include a damage mask matrix D (used to mark the location of failure elements). Step 2: Generation Stage Based on Hypernetwork: An improved Conditional Variational Autoencoder (CVAE) is used as the generator, and a hypernetwork module is introduced. The generator includes an encoder, a decoder, and a hypernetwork module. The hypernetwork module takes the geometric vector g as input and dynamically generates the weights and bias parameters of the decoder's hidden layers. The generator outputs a candidate structure pattern P1 based on the conditional information of the input sample pair (F,G). This stage improves the clarity and discreteness of the generated pattern through an adversarial learning mechanism (Adversarial loss). Adversarial learning is implemented through a discriminator, which is a 4-layer fully connected network. The input is the generated pattern and the geometric conditions. The training objective is to maximize the discrimination accuracy of real samples and minimize the discrimination accuracy of generated samples. Step 3: Optimization stage: A fully connected neural network is used as the optimizer. The optimizer is input to the geometric structure matrix G and the candidate structure pattern P1, and outputs the optimized structure P2. The optimizer establishes a far-field prediction model and adopts a greedy iterative strategy to scan the metasurface unit one by one to update the state until the error is lower than the preset threshold. Step 4: Output and Verification: Output the optimized metasurface structure P2 and verify whether its far-field distribution meets the target requirements through electromagnetic simulation or actual measurement. Verify whether the far-field distribution of P2 meets the target requirements through electromagnetic simulation tools (such as conventional HFSS, CST, etc.) or actual measurement. If it does not meet the requirements, return to Step 3 to adjust the iteration parameters and re-optimize.

[0009] In step one, the geometric structure matrix G has 20×20 dimensions, and the target far-field distribution matrix F has 30×30 dimensions.

[0010] Step two specifically includes: Flatten the target far-field matrix F and the geometric matrix G into vectors, and concatenate them into a conditional vector. [f,g]; The encoder encodes the input x into a latent variable z, where x , m is the vector after the complete pattern matrix M is flattened; The decoder generates a structural pattern based on [z,c], and its hidden layer parameters are output by the hypernetwork: in For hypernetwork mapping relationships, For hypernetic network parameters, This is the weight matrix. For bias, For activation functions; The generator loss function consists of three parts: in, Mean square error, To combat the losses, This represents the loss due to the maximum mean difference.

[0011] In step two, the encoder is a 5-layer fully connected neural network that maps the input far-field distribution matrix and geometric matrix to the latent space; the decoder is a 5-layer fully connected network that recovers the metasurface structure pattern from the latent variables; and the hypernetwork module is a 4-layer fully connected network used to dynamically generate parameters of a certain hidden layer of the decoder.

[0012] The optimizer in step three is a 3-layer fully connected neural network, including: Geometric feature extraction module: Extracts spatial features of the damaged area; Structural feature extraction module: Analyzes the unit distribution of the initial pattern; Fusion and Prediction Module: Predicts far-field response and evaluates error.

[0013] In step two, the adversarial learning mechanism is implemented through a discriminator, which is a 4-layer fully connected network. The input is the generated pattern and geometric conditions, and the training objective is to maximize the discrimination accuracy of real samples and minimize the discrimination accuracy of generated samples.

[0014] Step three specifically includes: the optimizer establishing a far-field prediction model S:[P,G]→F; Candidate structural patterns are calculated through forward prediction. Corresponding far-field distribution Far-field distribution of the target error ; A greedy iterative approach is used to attempt state switching for each metasurface unit. The error change ΔE after the switch is calculated. If ΔE < 0, the modification is retained; otherwise, the original state is restored. The iteration is repeated until the error E is lower than the preset threshold or the maximum number of iterations is reached.

[0015] The geometric damage is one or more of the following: metasurface structure breakage, unit missingness, corrosion, and shading.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a "Generate-Optimize (GO)" framework, combined with a hypernetwork dynamic parameter adjustment mechanism. It is implemented through four steps: input data preparation, a hypernetwork-based generation stage, an optimization stage, and output verification. The generation stage uses an improved CVAE as the generator, with the hypernetwork dynamically generating decoder parameters and incorporating adversarial learning and MMD constraints. The optimization stage uses a fully connected neural network optimizer to iteratively refine candidate patterns through greedy algorithms. This invention solves the problems of poor robustness, output ambiguity, and insufficient real-time performance in traditional inversion design when facing geometric damage conditions, exhibiting high robustness, high definition, high real-time performance, high accuracy, and strong adaptability.

[0017] High robustness: The super network introduces a geometric perception mechanism, enabling the model to adaptively identify geometric damage information. It can still accurately predict the metasurface structure in scenarios such as structural damage, missing units, corrosion or occlusion, without the need for retraining for specific damage. High definition: Combining adversarial learning and MMD constraints, the problem of fuzzy output in traditional models is effectively eliminated, enabling precise discretization control of metasurface structures with a pattern sharpness index of 0.87; High real-time performance: The overall inference time of the generation-optimization framework is less than 1ms, of which the generation stage takes less than 0.5ms, which can meet the requirements of real-time electromagnetic control scenarios; High accuracy: Experimental results show that, compared with the traditional DNN model, the mean far-field error (MSE) of this method is reduced from 0.122 to 0.021, a reduction of about 82.8%, and the Spearman correlation coefficient is increased from 0.354 to 0.692, an improvement of about 2 times; Strong adaptability: The same model can be compatible with multiple geometric damage types without adjusting the network structure or retraining, significantly reducing application costs; High optimization efficiency: Compared to random initialization optimization which requires 69 iterations, the generative-optimization framework of this invention can converge in only 20-30 iterations, reducing the average error from 73.6% to 35.5%. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the overall framework of the global inversion design method for geometrically damaged metasurfaces based on super-learning according to the present invention.

[0020] Figure 2 This is a flowchart of the dynamic weight generation process of the supernetwork in the decoder for a global inversion design method for geometrically damaged metasurfaces based on super learning, according to the present invention.

[0021] Figure 3 This is a flowchart of the optimizer iterative correction process for a global inversion design method for geometrically damaged metasurfaces based on super-learning, according to the present invention.

[0022] Figure 4 This is a comparison diagram of random optimization and GO model optimization for a global inversion design method for geometrically damaged metasurfaces based on super learning, as proposed in this invention.

[0023] Figure 5 The image shows the average prediction error of DNN, CNN, single generator, and GO models for a global inversion design method for geometrically damaged metasurfaces based on super learning, as presented in this invention.

[0024] Figure 6 This presents the experimental results of far-field prediction for a global inversion design method for geometrically damaged metasurfaces based on super-learning, as described in this invention. The blue curve represents the desired target, while the red and orange curves represent the far fields of the GO model and the DNN, respectively. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Reference Figure 1 A global inversion design method for geometrically damaged metasurfaces based on super-learning includes the following steps: Step 1: Input Data Preparation: Obtain a dataset containing the metasurface geometry matrix, the target far-field distribution matrix, and structural damage information to form an input sample pair (F,G) and the real structural pattern P; the input data may include a damage mask matrix D (used to mark the location of failure elements); the geometry matrix G is 20×20 in dimensions, and the target far-field distribution matrix F is 30×30 in dimensions.

[0027] Step 2: Generation Stage Based on Hypernetwork: An improved Conditional Variational Autoencoder (CVAE) is used as the generator, and a hypernetwork module is introduced. The generator includes an encoder, a decoder, and a hypernetwork module. The hypernetwork module takes the geometric vector g as input and dynamically generates the weights and bias parameters of the decoder's hidden layers. The generator outputs a candidate structure pattern P1 based on the conditional information of the input sample pair (F,G). This stage improves the clarity and discreteness of the generated pattern through an adversarial learning mechanism (Adversarial loss). Adversarial learning is implemented through a discriminator, which is a 4-layer fully connected network. The input is the generated pattern and the geometric conditions. The training objective is to maximize the discrimination accuracy of real samples and minimize the discrimination accuracy of generated samples. Generator: Used to quickly generate candidate metasurface patterns under given geometry and far-field conditions.

[0028] The concrete implementation of the generator: (1) Network composition: The generator adopts a conditional variational autoencoder (CVAE) structure, which mainly includes: Encoder: A 5-layer fully connected neural network that maps the input far-field distribution matrix and geometric matrix to the latent space.

[0029] Decoder: A 5-layer fully connected network that recovers metasurface structure patterns from latent variables.

[0030] Hypernetwork module: A 4-layer fully connected network used to dynamically generate parameters for a hidden layer of the decoder.

[0031] Input data includes: geometric matrix Target far-field matrix ; An optional damage mask matrix D is used to mark the location of the failed element.

[0032] Flatten all matrices and concatenate them into a vector: Encoder output latent variable Then, sampling is performed using reparameterization techniques.

[0033] (2) Hypernetwork parameter generation mechanism In this invention, the input to the hypernetwork is a geometric vector. The output is a set of dynamic : in: : The weight matrix corresponding to a certain layer of the decoder.

[0034] Bias term These parameters will replace the original fixed parameters during the inference phase and be used to dynamically adjust the decoder's structure mapping: Where D represents a dynamic network.

[0035] This design enables the network to adaptively adjust its generation pattern based on current geometric changes, significantly improving its robustness to structural damage.

[0036] (3) Training objective function: The generator's total loss consists of three parts: in: Forward fitting loss constrains the generated output to match the real pattern.

[0037] To combat loss, the clarity of the pattern is enhanced through a discriminator.

[0038] Maximum mean difference loss is used to maintain a smooth and stable latent space distribution.

[0039] Weight parameters .

[0040] (4) Counter-structural design To reduce blurry output, a discriminator is connected after the decoder, whose input is the generated pattern and geometric conditions. The discriminator is a 4-layer fully connected network, with the following training objectives: Through this game-theoretic optimization mechanism, the generator gradually outputs more realistic discrete structures.

[0041] (5) Output and performance indicators Candidate patterns output by the generator It has the following characteristics: Generation time < 0.5ms; Pattern sharpness index ,in This represents the total number of primary atoms on a single metasurface. This indicates the number of high-quality elementary atoms; It can be directly used as input to the optimizer for subsequent fine-tuning.

[0042] Step 3: Optimization Stage: A fully connected neural network is used as the optimizer. The optimizer takes the geometric structure matrix G and the candidate structure pattern P1 as input and outputs the optimized structure P2. The optimizer establishes a far-field prediction model and uses a greedy iterative strategy to scan each metasurface unit and update its state until the error is lower than a preset threshold. Optimizer: Efficiently refines and corrects the generated pattern to further improve the matching accuracy with the target far field.

[0043] Specific implementation of the optimizer (1) Structure of the optimizer The optimizer is a 3-layer fully connected neural network, including: Geometric feature extraction module: Extracts spatial features of the damaged area; Structural feature extraction module: Analyzes the unit distribution of the initial pattern; Fusion and Prediction Module: Predicts far-field response and evaluates error.

[0044] The optimizer takes (P1, G) as input and outputs the corrected pattern P2.

[0045] (2) Iterative optimization strategy The optimizer uses a greedy iterative algorithm based on the generated initial pattern: 2.1 Perform a state switching attempt for each meta-atom; 2.2 Calculate the error change ΔE in the far field prediction after the switch; 2.3 If ΔE < 0, then retain this modification; 2.4 Repeat the iteration until the convergence condition is met or the maximum number of iterations is reached.

[0046] In experiments, the error typically converges after 20-30 iterations.

[0047] (3) Optimize performance Through the above mechanism, the optimizer can make rapid corrections without changing the network parameters.

[0048] Comparative experiments show that: Random initialization requires 69 iterations; Using the generative-optimization framework of this invention requires only 26 iterations; The average error decreased from 73.6% to 35.5%.

[0049] Step 4: Output and Verification: Output the optimized metasurface structure P2 and verify whether its far-field distribution meets the target requirements through electromagnetic simulation or actual measurement. Verify whether the far-field distribution of P2 meets the target requirements through electromagnetic simulation tools (such as conventional HFSS, CST, etc.) or actual measurement. If it does not meet the requirements, return to Step 3 to adjust the iteration parameters and re-optimize.

[0050] Step two specifically includes: Flatten the target far-field matrix F and the geometric matrix G into vectors, and concatenate them into a conditional vector. [F,G]; The encoder encodes the input x into a latent variable z; The decoder generates a structural pattern based on [z,c], and its hidden layer parameters are output by the hypernetwork: in For hypernetwork mapping relationships, For hypernetic network parameters, This is the weight matrix. For bias, For activation functions; The generator loss function consists of three parts: in, Mean square error, To combat the losses, This represents the loss due to the maximum mean difference.

[0051] In step two, the encoder is a 5-layer fully connected neural network that maps the input far-field distribution matrix and geometric matrix to the latent space; the decoder is a 5-layer fully connected network that recovers the metasurface structure pattern from the latent variables; and the hypernetwork module is a 4-layer fully connected network used to dynamically generate parameters of a certain hidden layer of the decoder.

[0052] The optimizer in step three is a 3-layer fully connected neural network, including: Geometric feature extraction module: Extracts spatial features of the damaged area; Structural feature extraction module: Analyzes the unit distribution of the initial pattern; Fusion and Prediction Module: Predicts far-field response and evaluates error.

[0053] In step two, the adversarial learning mechanism is implemented through a discriminator, which is a 4-layer fully connected network. The input is the generated pattern and geometric conditions, and the training objective is to maximize the discrimination accuracy of real samples and minimize the discrimination accuracy of generated samples.

[0054] Step three specifically includes: the optimizer establishing a far-field prediction model S:[P,G]→F; Candidate structural patterns are calculated through forward prediction. Corresponding far-field distribution Far-field distribution of the target error ; A greedy iterative approach is used to attempt state switching for each metasurface unit. The error change ΔE after the switch is calculated. If ΔE < 0, the modification is retained; otherwise, the original state is restored. The iteration is repeated until the error E is lower than the preset threshold or the maximum number of iterations is reached.

[0055] The geometric damage is one or more of the following: metasurface structure breakage, unit missingness, corrosion, and shading.

[0056] Experiments and verification: Under 5GHz conditions, using Simulation and experimental measurement of the two-phase unit of the symmetrical metal ring.

[0057] The model was trained using 70,000 samples with different geometric damage patterns, and the validation results are as follows:

[0058] Experimental results show that the present invention can maintain high precision and high real-time performance under various geometric damage scenarios, which is significantly better than traditional methods.

[0059] Figure 4 (a) Explanation: To verify the performance of the GO model in terms of pattern prediction accuracy and real-time performance, this invention selects a randomly generated four-beam far-field as the test target and sets up two sets of control methods: one set uses the output results of the GO model proposed in this invention as the initial optimization value, and the other set uses a random mode for initialization, and performs optimization iterations respectively. To quantitatively evaluate the prediction performance, this invention defines the average prediction error as: in To predict far-field data for metasurface patterns, For the target far-field data, N is the number of test samples.

[0060] Although both methods achieve similar prediction accuracy after iterative convergence, the GO model of this invention has a significant advantage in the initial stage: its initial prediction error is 35.5, significantly lower than the 73.6 of the random initialization method. Furthermore, the GO model requires only 26 effective iterations to converge to the target far field, while the random method requires 69 iterations.

[0061] Figure 4(b) Explanation: For the random initialization method, the initial unit arrangement is completely disordered. At iterations t=1, t=17, and t=34, the pattern structure still exhibits a highly random distribution, only gradually approaching the structural characteristics of the target pattern at t=69. The corresponding far-field distribution shows significant deviations in the early iterations, with the main lobe direction and amplitude differing significantly from the target far-field. The far-field pattern exhibits blurry and distorted characteristics, only gradually approaching the target response after numerous iterations. In contrast, the initial pattern generated using the GO model exhibits obvious structural characteristics from the beginning of the iterations. For example, a relatively clear pattern outline is present at iteration t=1; the pattern further converges towards the target mode at iterations t=6 and t=13; and it basically reaches a stable state at iteration t=26. Its corresponding far-field distribution has a high similarity to the target far-field in the initial stage. As the number of iterations increases, the shape, direction, and energy distribution of the far-field main lobe gradually match the target, with a convergence speed significantly faster than the random initialization method.

[0062] Figure 5 To more comprehensively verify the predictive performance of the GO model, the test dataset was uniformly divided into four groups (T1 to T4), each containing 750 test samples. Subsequently, Convolutional Neural Networks (CNN), Fully Connected Neural Networks (DNN), a single generator model, and the GO model proposed in this specification were used to predict the results of these four groups, and the average error for each group was calculated. Statistical analysis of the prediction results of the four models shows that the prediction error decreases progressively from DNN to the GO model. Specifically, the average errors of DNN and CNN are 57.6 and 55.7, respectively, with little difference, reflecting their consistency in network structure and training methods. When using the single generator model, the average error is reduced to 45.6, approximately 20% lower than the previous two types of networks. The GO model proposed in this invention has the lowest error among the four methods, with an average error of 35.6, approximately 40% lower than both CNN and DNN.

[0063] Figure 6 To verify the predictive capability of the GO model in actual metasurface structures, experimental tests were conducted on the far-field of the designed four-beam target, and the model's prediction results were compared and analyzed with those of a standard DNN model. In the experimental setup, four typical directions were selected as verification targets, with two beams located at the azimuth angle. Pitch angle and Direction, the other two beams are located Pitch angle and Direction. For each target direction, several different geometric damage structures are randomly generated to simulate real-world application scenarios under complex deployment conditions. For example... Figure 6 As shown, the blue curve represents the theoretical target's far field, the red curve represents the prediction result of the GO model, and the yellow curve represents the prediction result of the DNN model. The upper part of the figure corresponds to the azimuth angle. The experimental results, the second half of the results correspond to The experimental results. From Figure 6 As can be seen, the GO model can accurately predict the target beam direction in all test samples, and the angular positions of each main lobe are highly consistent with the target distribution. The far-field amplitude distribution also maintains a good match, demonstrating strong robustness and accuracy. In contrast, the prediction results of the DNN model show significant deviations. Its far field usually exhibits a single-sided main lobe and high-amplitude specular reflection components, and the main lobe direction also shows varying degrees of shift, failing to match the target's far field.

[0064] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A global inversion design method for geometrically damaged metasurfaces based on super-learning, characterized in that: Includes the following steps: Step 1: Input data preparation: Obtain a dataset containing the metasurface geometry matrix, the target far-field distribution matrix, and structural damage information to form input sample pairs (F,G) and the real structural pattern P; Step 2: Generation stage based on supernetwork: An improved conditional variational autoencoder (CVAE) is used as the generator. The generator includes an encoder, a decoder, and a supernetwork module. The supernetwork module takes the geometric vector g as input and dynamically generates the weights and bias parameters of the decoder's hidden layer. The generator outputs a candidate structure pattern P1 based on the conditional information of the input sample pair (F,G). This stage improves the clarity and discreteness of the generated pattern through an adversarial learning mechanism. Step 3: Optimization stage: A fully connected neural network is used as the optimizer. The optimizer is input to the geometric structure matrix G and the candidate structure pattern P1, and outputs the optimized structure P2. The optimizer establishes a far-field prediction model and adopts a greedy iterative strategy to scan the metasurface unit one by one to update the state until the error is lower than the preset threshold. Step 4: Output and Verification: Output the optimized metasurface structure P2 and verify whether its far-field distribution meets the target requirements through electromagnetic simulation or actual measurement.

2. The global inversion design method for geometrically damaged metasurfaces based on super-learning according to claim 1, characterized in that: In step one, the geometric structure matrix G has 20×20 dimensions, and the target far-field distribution matrix F has 30×30 dimensions.

3. The global inversion design method for geometrically damaged metasurfaces based on super-learning according to claim 1, characterized in that: Step two specifically includes: Flatten the target far-field matrix F and the geometric matrix G into vectors, and concatenate them into a conditional vector. [f,g]; The encoder encodes the input x into a latent variable z; The decoder generates a structural pattern based on [z,c], and its hidden layer parameters are output by the hypernetwork: in This represents the hypernetwork mapping relationship. For hypernetic network parameters, This is the weight matrix. For bias, For activation functions; The generator loss function consists of three parts: in, Mean square error, To combat the losses, This represents the loss due to the maximum mean difference.

4. The global inversion design method for geometrically damaged metasurfaces based on super-learning according to claim 1, characterized in that: In step two, the encoder is a 5-layer fully connected neural network that maps the input far-field distribution matrix and geometric matrix to the latent space. The decoder is a 5-layer fully connected network that recovers metasurface structure patterns from latent variables; The hypernetwork module is a 4-layer fully connected network used to dynamically generate parameters for a hidden layer of the decoder.

5. The global inversion design method for geometrically damaged metasurfaces based on super-learning according to claim 1, characterized in that: The optimizer in step three is a 3-layer fully connected neural network, including: Geometric feature extraction module: Extracts spatial features of the damaged area; Structural feature extraction module: Analyzes the unit distribution of the initial pattern; Fusion and Prediction Module: Predicts far-field response and evaluates error.

6. The global inversion design method for geometrically damaged metasurfaces based on super-learning according to claim 1, characterized in that: In step two, the adversarial learning mechanism is implemented through a discriminator, which is a 4-layer fully connected network. The input is the generated pattern and geometric conditions, and the training objective is to maximize the discrimination accuracy of real samples and minimize the discrimination accuracy of generated samples.

7. The global inversion design method for geometrically damaged metasurfaces based on super-learning according to claim 5, characterized in that: Step three specifically includes: the optimizer establishing a far-field prediction model S:[P,G]→F; Candidate structural patterns are calculated through forward prediction. Corresponding far-field distribution Far-field distribution of the target error ; A greedy iterative approach is used to attempt state switching for each metasurface unit. The error change ΔE after the switch is calculated. If ΔE < 0, the modification is retained; otherwise, the original state is restored. The iteration is repeated until the error E is lower than the preset threshold or the maximum number of iterations is reached.

8. The global inversion design method for geometrically damaged metasurfaces based on super-learning according to claim 1, characterized in that: The geometric damage is one or more of the following: metasurface structure breakage, unit missingness, corrosion, and shading.

Images