A metasurface antenna parameter optimization method and system based on a convolutional neural network

Abstract

The application relates to the technical field of super surface antennas, and discloses a super surface antenna parameter optimization method and system based on a convolutional neural network, which comprises the following steps: collecting super surface sample data; extracting a comprehensive electromagnetic feature set and establishing a nonlinear mapping relationship model; constructing an antenna performance comprehensive evaluation function; inputting the nonlinear mapping relationship model and the antenna performance comprehensive evaluation function into a hybrid optimization algorithm to generate a parameter candidate set; using a particle swarm optimization algorithm to locally optimize the parameter candidate set to obtain a super surface antenna parameter combination; extracting electromagnetic features of the super surface antenna parameter combination; iteratively adjusting a resonant cavity height parameter until a threshold value is met to obtain electromagnetic simulation verification results; feeding back the electromagnetic simulation verification results to a deep Q neural network model for parameter updating to obtain an optimal super surface antenna parameter combination. The application realizes optimization of antenna parameters, improves the design efficiency of super surface antennas, and reduces the consumption of electromagnetic simulation calculation resources.

Description

Technical Field

[0001] This invention relates to the field of metasurface antenna technology, and in particular to a method and system for optimizing metasurface antenna parameters based on convolutional neural networks. Background Technology

[0002] Metasurface antennas are a novel type of antenna structure based on metamaterial technology. They control the propagation characteristics of electromagnetic waves by designing artificial microstructure units on specific surfaces. Compared with traditional antennas, metasurface antennas have advantages such as low profile, high gain, strong directivity, and multi-band operation capabilities, and have broad application prospects in modern wireless communication, radar detection, and the Internet of Things. The performance of metasurface antennas is highly dependent on the precise design of their structural parameters, including key parameters such as surface unit period, resonant cavity height, and metamaterial unit size. These parameters have complex electromagnetic coupling relationships that affect the antenna's core performance indicators such as operating frequency, bandwidth, gain, and impedance matching.

[0003] Existing technologies employ traditional data-driven methods (such as regression analysis and interpolation) combined with electromagnetic simulation software for parameter mapping, achieving basic modeling of the relationship between antenna parameters and electromagnetic performance. However, this method suffers from significant drawbacks when processing high-dimensional electromagnetic field data, including limited feature extraction capabilities, a lack of deep spatial feature learning, and insufficient nonlinear mapping capabilities. This leads to an inaccurate grasp of the complex relationship between resonant cavity height and antenna performance, and an inability to achieve intelligent mapping between electromagnetic field distribution information and optimization parameters. Particularly in multi-band scenarios, it is even more difficult to achieve adaptive optimization of antenna parameters and precise control of performance indicators, ultimately resulting in low efficiency, high resource consumption, and poor stability of optimization results in metasurface antenna design. Summary of the Invention

[0004] In view of this, the present invention proposes a method and system for optimizing metasurface antenna parameters based on convolutional neural networks. This method solves the significant shortcomings of existing technologies in processing high-dimensional electromagnetic field data, such as limited feature extraction capabilities, lack of deep spatial feature learning, and insufficient nonlinear mapping capabilities. These shortcomings lead to inaccurate understanding of the complex relationship between resonant cavity height and antenna performance, and an inability to achieve intelligent mapping between electromagnetic field distribution information and optimization parameters. Especially in multi-band scenarios, it is even more difficult to achieve adaptive optimization of antenna parameters and precise control of performance indicators, ultimately resulting in low efficiency, high resource consumption, and poor stability of optimization results in metasurface antenna design.

[0005] The technical solution of this invention is implemented as follows: In a first aspect, this invention provides a method for optimizing metasurface antenna parameters based on a convolutional neural network, comprising the following steps:

[0006] Raw sample data of the metasurface antenna is collected and preprocessed to obtain metasurface sample data;

[0007] A two-stream convolutional neural network model is constructed, and spatial domain features and frequency domain features are extracted from the metasurface sample data through the two-stream convolutional neural network model. A comprehensive electromagnetic feature set is constructed based on the spatial domain features and frequency domain features, and a nonlinear mapping relationship model between electromagnetic features and resonant cavity height is established.

[0008] A comprehensive evaluation function for antenna performance is constructed. The nonlinear mapping relationship model and the comprehensive evaluation function for antenna performance are input into a hybrid optimization algorithm. A parameter candidate set is generated by a genetic algorithm, and the parameter candidate set is locally optimized by a particle swarm optimization algorithm to obtain the combination of metasurface antenna parameters.

[0009] The electromagnetic characteristics of the metasurface antenna parameter combination are extracted. The resonant cavity height parameter is iteratively adjusted based on the matching degree of gain and impedance bandwidth until a threshold is met, and the electromagnetic simulation verification results are obtained. A parameter optimization closed loop is constructed through a reinforcement learning mechanism, and the electromagnetic simulation verification results are fed back to the deep Q neural network model for parameter update, so as to obtain the optimal metasurface antenna parameter combination.

[0010] Based on the above technical solutions, preferably, the step of acquiring raw sample data of the metasurface antenna and performing preprocessing to obtain metasurface sample data includes:

[0011] By simulating metasurface antennas with different structural parameters and resonant cavity heights using electromagnetic simulation, the corresponding electromagnetic field distribution maps and S-parameter curves are obtained, thus yielding the original sample data of the metasurface antennas.

[0012] The original sample dataset is standardized and enhanced to eliminate dimensional differences. Data quality is improved through data denoising, missing value processing, and image enhancement to construct a training metasurface sample dataset. The metasurface sample data includes antenna structure parameters, electromagnetic field distribution maps, and S-parameter curves.

[0013] Based on the above technical solutions, preferably, the construction of the dual-stream convolutional neural network model, extracting spatial domain features and frequency domain features from the metasurface sample data through the dual-stream convolutional neural network model, constructing a comprehensive electromagnetic feature set based on the spatial domain features and frequency domain features, and establishing a nonlinear mapping relationship model between electromagnetic features and resonant cavity height, includes:

[0014] A two-stream convolutional neural network is constructed, which includes a spatial domain feature extraction branch and a frequency domain feature extraction branch. The spatial domain feature extraction branch uses three-dimensional convolution to process the electromagnetic field distribution map, and the frequency domain feature extraction branch uses one-dimensional convolution to process the S-parameter curve. The features of the two branches are merged through a feature fusion layer to construct a comprehensive electromagnetic feature set.

[0015] A nonlinear mapping model is trained using the comprehensive electromagnetic feature set and the corresponding resonant cavity height data. The nonlinear mapping model adopts a multilayer perceptron structure.

[0016] Based on the above technical solutions, preferably, the construction of the dual-stream convolutional neural network includes a spatial domain feature extraction branch and a frequency domain feature extraction branch, wherein:

[0017] The spatial domain feature extraction branch employs a residual-structured 3D convolutional network consisting of three convolutional units. Each convolutional unit contains two 3×3×3 convolutional layers, a batch normalization layer, and a PReLU activation function, with residual connections. The number of convolutional kernels increases progressively from 32 to 128. The frequency domain feature extraction branch uses an attention-enhanced 1D convolutional network, including a temporal attention module and a frequency attention module. The frequency domain feature extraction branch uses multi-scale convolutional kernels to process S-parameter curves in parallel, capturing antenna response characteristics across different frequency ranges. The feature fusion layer employs an adaptive weighting mechanism to integrate features extracted from the spatial and frequency domain feature extraction branches, dynamically adjusting the weights of different features based on the importance of specific frequency bands.

[0018] Based on the above technical solutions, preferably, the construction of the antenna performance comprehensive evaluation function involves inputting the nonlinear mapping relationship model and the antenna performance comprehensive evaluation function into a hybrid optimization algorithm. A parameter candidate set is generated using a genetic algorithm, and the parameter candidate set is locally optimized using a particle swarm optimization algorithm to obtain the metasurface antenna parameter combination, including:

[0019] A comprehensive performance evaluation function for multi-target antennas is constructed. Adaptive weighting coefficients for frequency bands are introduced. Different evaluation standards are set for bandwidth, gain, directivity and impedance matching parameters of different frequency bands, and weighted calculations are performed to obtain the comprehensive performance evaluation function for antennas.

[0020] A hybrid optimization algorithm framework is established, which integrates the nonlinear mapping relationship model with the antenna performance comprehensive evaluation function. A genetic algorithm is used to generate a candidate set of metasurface antenna parameters. An adaptive mutation strategy is used to enhance the parameter space exploration capability of the candidate set of metasurface antenna parameters, resulting in several candidate parameters. The candidate parameters are then locally optimized based on the quantum particle swarm optimization algorithm to obtain the combination of metasurface antenna parameters.

[0021] Based on the above technical solutions, preferably, the constructed antenna performance comprehensive evaluation function includes:

[0022] To address the operating frequency band characteristics of metasurface antennas, the target frequency range is divided into multiple sub-bands, with different priorities assigned to each band. Normalized bandwidth, gain, directivity, and impedance matching metrics are defined for each band. Frequency band correlation analysis is used to identify coupling effects between bands, and a performance balance model between bands is established. For multi-band application scenarios, a hierarchical evaluation structure is constructed. Principal component analysis is used to reduce redundancy among evaluation metrics and extract key performance characteristics. Fuzzy comprehensive evaluation is employed to integrate the evaluation results of multiple bands and metrics, resulting in a single comprehensive antenna performance evaluation function.

[0023] The establishment of the hybrid optimization algorithm framework includes:

[0024] After initializing the population, an improved genetic algorithm with adaptive crossover and mutation strategies is used to generate a candidate set of antenna parameters. The crossover and mutation probabilities are dynamically adjusted according to the population diversity and fitness distribution. The Pareto dominance ranking mechanism is used to handle the multi-objective optimization problem, retaining the non-dominated solution set. For each candidate solution in the non-dominated solution set, a nonlinear mapping relationship model is used to predict the optimal resonant cavity height. The predicted optimal resonant cavity height is combined with the metasurface antenna parameters to obtain the complete solution. The quantum particle swarm optimization algorithm is used for local optimization, where the quantum rotation angle and contraction / expansion factor are adaptively adjusted based on the quality of the complete solution. A dynamic balance mechanism for solution space exploration and development is set up to adjust the ratio of global search to local refinement according to the optimization process.

[0025] Based on the above technical solutions, preferably, the extraction of electromagnetic characteristics of the metasurface antenna parameter combination, iterative adjustment of the resonant cavity height parameter based on the matching degree of gain and impedance bandwidth until a threshold is met, obtaining electromagnetic simulation verification results, constructing a parameter optimization closed loop through a reinforcement learning mechanism, and feeding the electromagnetic simulation verification results back to the deep Q neural network model for parameter update, obtaining the optimal metasurface antenna parameter combination, including:

[0026] A real-time electromagnetic field distribution feature extraction mechanism is constructed. Electromagnetic simulation is performed on the combination of metasurface antenna parameters generated by the hybrid optimization algorithm. Electromagnetic field distribution map and S-parameter curve are obtained in real time. Electromagnetic features are extracted through a pre-trained convolutional neural network model. An iterative adjustment strategy for the resonant cavity height parameter is constructed. The resonant cavity height is adjusted according to the gain and impedance bandwidth matching degree.

[0027] A parameter optimization closed loop based on deep reinforcement learning is constructed. The comprehensive evaluation function of antenna performance is used as the reward signal to establish a state-action-reward mapping relationship. The optimization strategy of metasurface antenna parameters is learned through deep Q neural network. The electromagnetic simulation verification results are fed back to the model weight update process to perform adaptive closed-loop optimization between metasurface antenna parameters and electromagnetic characteristics.

[0028] Secondly, the present invention also provides a metasurface antenna parameter optimization system based on a convolutional neural network, the system comprising:

[0029] The sample data acquisition module is used to collect raw sample data of the metasurface antenna and preprocess it to obtain metasurface sample data.

[0030] The mapping model building module is used to construct a two-stream convolutional neural network model. Through the two-stream convolutional neural network model, spatial domain features and frequency domain features are extracted from the metasurface sample data. Based on the spatial domain features and frequency domain features, a comprehensive electromagnetic feature set is constructed, and a nonlinear mapping relationship model between electromagnetic features and resonant cavity height is established.

[0031] The antenna parameter optimization module is used to construct a comprehensive evaluation function for antenna performance. The nonlinear mapping relationship model and the comprehensive evaluation function for antenna performance are input into a hybrid optimization algorithm. A parameter candidate set is generated through a genetic algorithm, and the parameter candidate set is locally optimized using a particle swarm optimization algorithm to obtain the combination of metasurface antenna parameters.

[0032] The simulation verification feedback module is used to extract the electromagnetic characteristics of the metasurface antenna parameter combination. Based on the matching degree of gain and impedance bandwidth, the resonant cavity height parameter is iteratively adjusted until a threshold is met to obtain the electromagnetic simulation verification result. A parameter optimization closed loop is constructed through a reinforcement learning mechanism, and the electromagnetic simulation verification result is fed back to the deep Q neural network model for parameter update to obtain the optimal metasurface antenna parameter combination.

[0033] Thirdly, the present invention also provides an electronic device, comprising: at least one processor, at least one memory, a communication interface, and a bus;

[0034] The processor, memory, and communication interface communicate with each other through the bus. The memory stores program instructions that can be executed by the processor. The processor calls the program instructions to implement the steps of a metasurface antenna parameter optimization method based on a convolutional neural network.

[0035] Fourthly, the present invention also provides a computer-readable storage medium storing computer instructions that enable a computer to implement steps such as those of a method for optimizing metasurface antenna parameters based on a convolutional neural network.

[0036] The method and system for optimizing metasurface antenna parameters based on convolutional neural networks of the present invention have the following advantages over the prior art:

[0037] (1) By using a dual-stream convolutional neural network, the electromagnetic features in the spatial and frequency domains are deeply extracted and fused, and an accurate nonlinear mapping relationship between electromagnetic features and resonant cavity height is established. Combined with a comprehensive evaluation function, a hybrid optimization algorithm and a reinforcement learning closed-loop mechanism, the antenna parameters are optimized, which significantly improves the design efficiency of metasurface antennas and reduces the consumption of electromagnetic simulation computing resources.

[0038] (2) Through the dual-stream convolutional neural network architecture and electromagnetic feature extraction mechanism, multi-scale spatial structure analysis of electromagnetic field distribution map and frequency domain feature mining of S-parameter curve were realized. The feature fusion strategy based on attention mechanism and multilayer perceptron were introduced to realize adaptive weighting and integration of feature saliency. The computational complexity was reduced by using deep separable convolution and adaptive pooling, which enhanced the ability to identify electromagnetic field anomaly patterns and realized a high-precision mapping relationship between electromagnetic features and resonant cavity height.

[0039] (3) By constructing an antenna performance comprehensive evaluation function and a hybrid optimization algorithm framework, integrating the improved genetic algorithm and quantum particle swarm optimization method, and combining the adaptive crossover mutation strategy, Pareto advantage sorting mechanism and simulated annealing strategy, the efficiency of local fine optimization is improved while ensuring the global optimal solution search capability. The dynamic relationship between exploration and development is balanced, the convergence speed and result stability of parameter optimization are improved, and the risk of getting trapped in local optima is reduced. Attached Figure Description

[0040] 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.

[0041] Figure 1 This is a flowchart of a metasurface antenna parameter optimization method based on a convolutional neural network according to the present invention.

[0042] Figure 2 This is a structural diagram of a metasurface antenna parameter optimization system based on a convolutional neural network according to the present invention. Detailed Implementation

[0043] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0044] Please see Figure 1 This invention provides a method for optimizing metasurface antenna parameters based on convolutional neural networks, comprising the following steps:

[0045] Raw sample data of the metasurface antenna is collected and preprocessed to obtain metasurface sample data, which includes antenna structural parameters, electromagnetic field distribution map and S-parameter curve.

[0046] A two-stream convolutional neural network model is constructed, and spatial domain features and frequency domain features are extracted from the metasurface sample data through the two-stream convolutional neural network model. A comprehensive electromagnetic feature set is constructed based on the spatial domain features and frequency domain features, and a nonlinear mapping relationship model between electromagnetic features and resonant cavity height is established.

[0047] A comprehensive evaluation function for antenna performance is constructed. The nonlinear mapping relationship model and the comprehensive evaluation function for antenna performance are input into a hybrid optimization algorithm. A parameter candidate set is generated by a genetic algorithm, and the parameter candidate set is locally optimized by a particle swarm optimization algorithm to obtain the combination of metasurface antenna parameters.

[0048] The electromagnetic characteristics of the metasurface antenna parameter combination are extracted. The resonant cavity height parameter is iteratively adjusted based on the matching degree of gain and impedance bandwidth until a threshold is met, and the electromagnetic simulation verification results are obtained. A parameter optimization closed loop is constructed through a reinforcement learning mechanism, and the electromagnetic simulation verification results are fed back to the deep Q neural network model for parameter update, so as to obtain the optimal metasurface antenna parameter combination.

[0049] Specifically, this embodiment uses a dual-stream convolutional neural network to deeply extract and fuse electromagnetic features in the spatial and frequency domains, establishes an accurate nonlinear mapping relationship between electromagnetic features and resonant cavity height, and combines a comprehensive evaluation function, a hybrid optimization algorithm, and a reinforcement learning closed-loop mechanism to optimize antenna parameters, significantly improving the design efficiency of metasurface antennas and reducing the consumption of electromagnetic simulation computational resources.

[0050] The process of acquiring raw sample data from the metasurface antenna and preprocessing it to obtain metasurface sample data includes:

[0051] By simulating metasurface antennas with different structural parameters and resonant cavity heights using electromagnetic simulation, the corresponding electromagnetic field distribution maps and S-parameter curves are obtained, thus yielding the original sample data of the metasurface antennas.

[0052] In one specific embodiment, the electromagnetic simulation process includes:

[0053] Within the preset parameter range, a combination of antenna structure parameters and resonant cavity height sample points is generated using the Latin hypercube sampling method to ensure that the samples are uniformly distributed in the parameter space. For each sample point, a three-dimensional model is established using full-wave electromagnetic simulation software, and its electromagnetic field distribution and scattering parameters are calculated. The electromagnetic field distribution map includes electric and magnetic field components in three orthogonal directions, and the S-parameter curve includes the amplitude and phase changes of the reflection coefficient and transmission coefficient within the target frequency range.

[0054] The original sample dataset is standardized and enhanced to eliminate dimensional differences. Data quality is improved through data denoising, missing value processing, and image enhancement to construct a training metasurface sample dataset.

[0055] In one specific embodiment, the data standardization and enhancement processing includes:

[0056] The antenna structure parameters are normalized to ensure that all parameter values ​​are within the [0,1] interval; Gaussian blurring and random rotation are applied to the electromagnetic field distribution map to increase sample diversity; sliding window averaging and wavelet denoising are used to remove high-frequency noise from the S-parameter curve, and the number of sampling points is unified through interpolation; according to the importance of the antenna's operating frequency band, different weights are applied to the data of different frequency bands to highlight the characteristic expression of key frequency bands.

[0057] Specifically, this embodiment obtains and optimizes high-quality training data by preprocessing the metasurface antenna sample data. This embodiment automatically extracts electromagnetic field distribution maps and S-parameter curves by combining the electromagnetic simulation software API interface, and applies methods such as wavelet transform denoising, Z-score normalization, and density clustering anomaly detection to ensure data quality. Furthermore, by introducing advanced data augmentation strategies such as geometric transformation, Gaussian noise perturbation, and adversarial sample generation, the effective sample set size is significantly expanded, solving key problems such as sample scarcity, data noise interference, and uneven distribution in metasurface antenna parameter optimization, and reducing the risk of overfitting.

[0058] The construction of a two-stream convolutional neural network model involves extracting spatial and frequency domain features from the metasurface sample data, constructing a comprehensive electromagnetic feature set based on these features, and establishing a nonlinear mapping model between the electromagnetic features and the resonant cavity height. This includes:

[0059] A two-stream convolutional neural network is constructed, which includes a spatial domain feature extraction branch and a frequency domain feature extraction branch. The spatial domain feature extraction branch uses three-dimensional convolution to process the electromagnetic field distribution map, and the frequency domain feature extraction branch uses one-dimensional convolution to process the S-parameter curve. The features of the two branches are merged through a feature fusion layer to construct a comprehensive electromagnetic feature set.

[0060] A nonlinear mapping model is trained using the comprehensive electromagnetic feature set and the corresponding resonant cavity height data. The nonlinear mapping model adopts a multilayer perceptron structure to realize the mapping relationship between electromagnetic features and the optimal resonant cavity height, and the generalization ability of the model is improved by regularization technology.

[0061] Specifically, this embodiment achieves multi-scale spatial structure analysis of electromagnetic field distribution maps and in-depth frequency domain feature mining of S-parameter curves through a dual-stream convolutional neural network architecture and electromagnetic feature extraction mechanism. It introduces a feature fusion strategy based on attention mechanism and multilayer perceptron to achieve adaptive weighting and integration of feature saliency. It adopts deep separable convolution and adaptive pooling to reduce computational complexity, enhance the ability to identify electromagnetic field anomaly patterns, and realize a high-precision mapping relationship between electromagnetic features and resonant cavity height.

[0062] The construction of the two-stream convolutional neural network includes a spatial domain feature extraction branch and a frequency domain feature extraction branch, wherein:

[0063] The spatial domain feature extraction branch employs a residual-structured 3D convolutional network consisting of three convolutional units. Each convolutional unit contains two 3×3×3 convolutional layers, a batch normalization layer, and a PReLU activation function, with residual connections. The number of convolutional kernels increases layer by layer from 32 to 128, enabling the extraction of multi-scale features from the electromagnetic field distribution map. The frequency domain feature extraction branch uses an attention-enhanced 1D convolutional network, including a temporal attention module and a frequency attention module. The frequency domain feature extraction branch uses multi-scale convolutional kernels to process S-parameter curves in parallel, capturing antenna response characteristics across different frequency ranges. The feature fusion layer employs an adaptive weighting mechanism to integrate features extracted from the spatial and frequency domain feature extraction branches, dynamically adjusting the weights of different features based on the importance of specific frequency bands.

[0064] In one specific embodiment, the training of the nonlinear mapping model includes:

[0065] Using a comprehensive electromagnetic feature set as input and the optimal resonant cavity height as output, a regression model with a multilayer perceptron structure is constructed. This model contains three fully connected layers with 512 and 256 hidden layer units, respectively, and employs a Swish activation function with Dropout. Model training utilizes an improved Adam optimizer and an adaptive learning rate scheduling strategy, and introduces a Bayesian hyperparameter optimization method based on Gaussian processes to automatically adjust model parameters. The loss function combines mean squared error and correlation loss, comprehensively considering the consistency of prediction accuracy and prediction trend. Five-fold cross-validation is used to evaluate model performance, ensuring the stability and accuracy of model predictions under different antenna structure parameters.

[0066] In one specific embodiment, the loss function of the nonlinear mapping relationship model is:

[0067] ;

[0068] in, Mean square error, To predict height With true height Pearson correlation coefficient, For model parameters The regularization term, for and The Weserstein distance in frequency response, , , These are the mean squared error, Pearson correlation coefficient, and weighting coefficients of the model parameters, respectively. It is an exponential function.

[0069] In one specific embodiment, the comparison between the dual-stream convolutional neural network structure of this embodiment and the traditional convolutional neural network is shown in Tables 1-3 below.

[0070] Table 1 Comparison of computational parameters for two-stream convolutional neural network structures

[0071] Network structure types Number of parameters (in ten thousand) Parameter reduction rate compared to traditional CNN Traditional convolutional neural networks 542.8 benchmark value This invention relates to spatial domain branching (depth-separating convolution). 92.3 83.0% This invention relates to a frequency domain branch (1D-CNN+LSTM). 98.4 81.9% The total number of dual-stream convolutional neural network structures in this invention 190.7 64.9%

[0072] Table 1 illustrates the advantages of the dual-stream convolutional neural network architecture in this embodiment in terms of computational resource optimization. By introducing deep separable convolution technology, the number of parameters in the spatial domain branch is reduced sharply from 5.428 million in the traditional convolutional neural network to 923,000, achieving a significant reduction of 83.0%. At the same time, the frequency domain branch adopts a hybrid architecture of one-dimensional convolutional neural network and long short-term memory network, which also reduces the number of parameters by 81.9%. The total number of parameters in the overall dual-stream fusion architecture is only 1.907 million, saving 64.9% of computational resources compared with the traditional method. This not only significantly reduces hardware resource requirements and power consumption, but also improves the deployment flexibility of the model on edge devices, providing key technical support for the lightweight implementation of metasurface antenna parameter optimization systems.

[0073] Table 2 Comparison of Electromagnetic Feature Extraction Performance

[0074] Evaluation indicators Traditional CNN methods This invention's dual-stream CNN method Performance improvement Resonant cavity height prediction accuracy 89.2% 97.4% 9.2% Electromagnetic field anomaly pattern recognition sensitivity 62.5% 89.4% 43.0% Feature extraction time (ms / sample) 145 58 60.0% Number of iterations required for model convergence 850 320 62.4%

[0075] Table 2 clearly presents the comprehensive advantages of the dual-stream convolutional neural network architecture in this embodiment in several key performance indicators of electromagnetic feature extraction. Among them, the accuracy of resonant cavity height prediction is improved to 97.4%, which is 9.2 percentage points higher than the traditional method. In particular, the sensitivity of electromagnetic field anomaly pattern recognition is significantly improved from 62.5% to 89.4%, achieving a significant improvement of 43.0%, which is crucial for capturing changes in electromagnetic properties caused by minute structural changes. At the same time, the feature extraction time is shortened to 58 milliseconds / sample, which is 60.0% faster, and the number of iterations required for model convergence is reduced from 850 to 320, which is 62.4% less. These performance improvements fully demonstrate the excellent performance of this embodiment in terms of electromagnetic feature extraction efficiency, accuracy and robustness, laying a solid foundation for the accurate optimization of metasurface antenna parameters.

[0076] Table 3 Analysis of the Adaptive Fusion Effect of Electromagnetic Features

[0077] Fusion Strategy Resonant cavity height prediction accuracy Inference speed (samples / second) Simple feature splicing 91.5% 48.3 Weighted average fusion 93.8% 46.7 This invention integrates attention mechanisms. 97.4% 42.1

[0078] Table 3 compares and analyzes the effects of three different feature fusion strategies, revealing the superiority of the attention mechanism fusion method in this embodiment: compared with simple feature splicing and weighted average fusion, the attention mechanism fusion in this embodiment achieves a significant advantage in the accuracy of resonant cavity height prediction, reaching a high accuracy of 97.4%, which is 5.9 percentage points higher than the simple splicing method; although the inference speed is slightly reduced (42.1 samples / second), this small performance cost is exchanged for a significant improvement in accuracy, proving that the adaptive feature attention mechanism in this embodiment can effectively identify and highlight the electromagnetic features most relevant to the prediction task, adjust the importance weights of different features, and achieve optimal feature representation and information utilization efficiency.

[0079] Specifically, this embodiment achieves multi-dimensional deep analysis of electromagnetic field data through a deep learning architecture: In the spatial domain branch, a deep separable convolutional structure with residual connections is used in conjunction with a multi-scale feature pyramid network to effectively capture local textures, global gradients, and phase change patterns in the electromagnetic field distribution map, while reducing the number of computational parameters by 83%; In the frequency domain branch, a hybrid architecture combining one-dimensional convolution and long short-term memory networks is introduced to accurately extract frequency response features, resonant point positions, and bandwidth change trends in the S-parameter curve; By introducing an adaptive feature attention mechanism, dynamic weight fusion of features from the two branches is achieved, significantly improving the model's ability to identify highly sensitive features of the resonant cavity; This embodiment effectively solves the problems of poor generalization and high computational complexity in electromagnetic feature extraction, reducing model parameters by 65% ​​while maintaining a prediction accuracy of 97.4%, and particularly improving sensitivity by 43% compared to traditional methods in electromagnetic field anomaly pattern recognition, laying a solid foundation for accurate mapping and optimization of metasurface antenna parameters.

[0080] The process involves constructing a comprehensive antenna performance evaluation function, inputting the nonlinear mapping model and the comprehensive antenna performance evaluation function into a hybrid optimization algorithm, generating a parameter candidate set through a genetic algorithm, and then using a particle swarm optimization algorithm to locally optimize the parameter candidate set to obtain the metasurface antenna parameter combination, including:

[0081] A comprehensive performance evaluation function for multi-target antennas is constructed. Adaptive weighting coefficients for frequency bands are introduced. Different evaluation standards are set for bandwidth, gain, directivity and impedance matching parameters of different frequency bands and weighted calculations are performed to obtain the comprehensive antenna performance evaluation function, and the overall performance of multi-band antennas is optimized.

[0082] A hybrid optimization algorithm framework is established, which integrates the nonlinear mapping relationship model with the antenna performance comprehensive evaluation function. A genetic algorithm is used to generate a candidate set of metasurface antenna parameters. An adaptive mutation strategy is used to enhance the parameter space exploration capability of the candidate set of metasurface antenna parameters, resulting in several candidate parameters. The quantum particle swarm optimization algorithm is used to perform local optimization on the candidate parameters to obtain a combination of metasurface antenna parameters, achieving a balance between global and local optima.

[0083] Specifically, this embodiment constructs an antenna performance comprehensive evaluation function and a hybrid optimization algorithm framework, integrates an improved genetic algorithm and a quantum particle swarm optimization method, and combines an adaptive crossover mutation strategy, a Pareto advantage sorting mechanism, and a simulated annealing strategy. This ensures the ability to search for the global optimal solution while improving the efficiency of local fine optimization, balances the dynamic relationship between exploration and development, improves the convergence speed and result stability of parameter optimization, and reduces the risk of getting trapped in local optima.

[0084] The constructed antenna performance comprehensive evaluation function includes:

[0085] To address the operating frequency characteristics of metasurface antennas, the target frequency range is divided into multiple sub-bands, with different priorities assigned to each band. Normalized bandwidth, gain, directivity, and impedance matching metrics are defined for each band. Frequency band correlation analysis is used to identify coupling effects between bands, and a performance balance model between bands is established. For multi-band application scenarios, a hierarchical evaluation structure is constructed. Principal component analysis is used to reduce redundancy among evaluation metrics and extract key performance characteristics. Fuzzy comprehensive evaluation is employed to integrate the evaluation results from multiple bands and metrics, resulting in a single comprehensive antenna performance evaluation function, achieving a unified quantitative evaluation of multi-target performance.

[0086] The establishment of the hybrid optimization algorithm framework includes:

[0087] After initializing the population, an improved genetic algorithm with adaptive crossover and mutation strategies is used to generate a candidate set of antenna parameters. The crossover and mutation probabilities are dynamically adjusted according to the population diversity and fitness distribution. The Pareto dominance ranking mechanism is used to handle the multi-objective optimization problem, retaining the non-dominated solution set. For each candidate solution in the non-dominated solution set, a nonlinear mapping relationship model is used to predict the optimal resonant cavity height. The predicted optimal resonant cavity height is combined with the metasurface antenna parameters to obtain the complete solution. The quantum particle swarm optimization algorithm is used for local optimization, where the quantum rotation angle and contraction / expansion factor are adaptively adjusted based on the quality of the complete solution. The convergence process of the algorithm is controlled by a simulated annealing strategy, allowing the algorithm to accept poor solutions in the early stages to escape local optima. A dynamic balance mechanism for solution space exploration and development is set up to adjust the ratio of global search to local refinement according to the optimization process.

[0088] In one specific embodiment, the adaptive mutation probability of the hybrid optimization algorithm is calculated as follows:

[0089] ;

[0090] ;

[0091] in, The mutation probability of the current iteration. and These are the lower and upper bounds of the mutation probability, respectively. This represents the current fitness value of the individual. and These are the maximum and minimum fitness values ​​in the current population, respectively. As an indicator of the current population diversity, To preset the diversity threshold, To control the sensitivity of mutation probability to fitness, Let be the Euclidean distance between the i-th individual and the j-th individual. Let be the weight function based on the dominance relationship between the i-th individual and the j-th individual.

[0092] In one specific embodiment, the comparison between the hybrid optimization algorithm of this embodiment and the traditional genetic algorithm and particle swarm optimization algorithm is shown in Table 4-7 below.

[0093] Table 4 Comparative Analysis of Search Capabilities of Hybrid Optimization Algorithms

[0094] Evaluation indicators Traditional Particle Swarm Optimization Algorithm Traditional genetic algorithm This invention provides a hybrid optimization algorithm framework. Average improvement rate compared to traditional methods Volume of solution space explored in a single iteration (relative value) 1.0 1.3 3.5 250% Global optimal solution discovery probability (%) 58.4 61.7 85.7 42.7% Convergence stability (standard deviation) 0.087 0.065 0.023 69.7% Search diversity maintenance cycle (number of iterations) 145 168 412 163.1%

[0095] Table 4 comprehensively showcases the groundbreaking progress of the hybrid optimization algorithm in search capability in this embodiment. Through a quantum encoding strategy, the solution space exploration volume in a single iteration is 3.5 times that of traditional methods, equivalent to a 250% increase in search breadth. The probability of discovering the global optimum increases from approximately 60% in traditional methods to 85.7%, an improvement of 42.7%, indicating that the hybrid optimization algorithm in this embodiment can more reliably find the true global optimum. The convergence stability standard deviation is reduced to 0.023, an average improvement of 69.7% compared to traditional methods, indicating a significant improvement in the consistency and repeatability of optimization results. Particularly noteworthy is the extension of the search diversity maintenance period to 412 iterations, more than 2.6 times that of traditional methods. This ensures that the hybrid optimization algorithm maintains efficient global search capability during long-term iterations, effectively avoiding the risk of early local convergence and providing a strong foundation for antenna optimization in complex parameter spaces.

[0096] Table 5 Comparison of Algorithm Convergence Efficiency and Computational Resource Consumption

[0097] Evaluation indicators Traditional Particle Swarm Optimization Algorithm Traditional genetic algorithm This invention provides a hybrid optimization algorithm framework. Improvement effect Average number of iterations required for convergence 876 743 338 58.2%↓ Time (s) required to converge to 90% of the optimal solution 284 312 108 63.8%↓ Performance degradation rate when parameter dimensions increase by 100% 47.3% 52.6% 18.4% 63.6%↓ CPU utilization (%) 87.5 92.3 76.8 14.3%↓

[0098] Table 5 presents the significant advantages of the hybrid optimization algorithm in terms of computational efficiency in this embodiment. The average number of iterations required for convergence is reduced significantly from about 800 in the traditional method to 338, saving 58.2% of the iteration computation. The time required to converge to 90% of the optimal solution is shortened from nearly 5 minutes to only 108 seconds, a speedup of 63.8%, which greatly accelerates the optimization process. Of particular note is the parameter dimension sensitivity index. When the parameter dimension increases by 100%, the performance reduction rate of the hybrid optimization algorithm in this embodiment is only 18.4%, which is 63.6% lower than the average of the traditional method. This proves that the hybrid optimization algorithm has excellent adaptability to high-dimensional parameter spaces. At the same time, while maintaining high performance, the CPU utilization rate is reduced by 14.3%, indicating that the hybrid optimization algorithm has also achieved a qualitative leap in the efficiency of computational resource utilization, providing an efficient and low-consumption computational solution for large-scale antenna parameter optimization tasks.

[0099] Table 6 Comparison of Multi-Band Performance Optimization Results for Metasurface Antennas

[0100] Performance indicators Traditional single algorithm optimization results Optimization results of the hybrid optimization algorithm framework of this invention Performance improvement Antenna gain (dBi) 8.45 10.62 25.7% Impedance bandwidth (MHz) 325 428 31.7% Return loss (dB) -18.7 -24.3 29.9% Multi-band performance balance (%) 72.4 91.8 26.8% Optimize overall performance indicators 0.649 0.828 27.5% Local optimal trap escape success rate (%) 43.8 89.5 104.3%

[0101] Table 6 provides a detailed comparison of the optimization results of the hybrid optimization algorithm framework of this embodiment and the traditional single algorithm in terms of the actual performance indicators of the metasurface antenna: the antenna gain is increased to 10.62 dBi, an improvement of 25.7%; the impedance bandwidth is extended to 428 MHz, an increase of 31.7%; the return loss is improved from -18.7 dB to -24.3 dB, an improvement of 29.9%; most importantly, the multi-band performance balance is improved from 72.4% to 91.8%, solving the problem that traditional methods are difficult to balance multi-band performance; the overall performance index is improved by 27.5%, reaching a high level of 0.828; the local optimal trap escape success rate is doubled to 89.5%, which means that the hybrid optimization algorithm framework of this embodiment can effectively avoid suboptimal solutions in complex nonlinear optimization landscapes and greatly improve the probability of finding the true global optimal solution.

[0102] Table 7. Analysis of the effectiveness of adaptive parameter adjustment strategy

[0103] Algorithm stage Explore weights Development weight Parameter update step size Convergence speed (relative value) Initial phase (1-20% iterations) 0.85 0.15 0.12 1.0 Mid-term stage (21-60% iterations) 0.58 0.42 0.08 2.4 Later stage (61-100% iterations) 0.23 0.77 0.03 3.7 Simulated annealing dynamic adjustment effect 0.45 (average) 0.55 (average) Adaptive 2.8 (average)

[0104] Table 7 reveals the dynamic evolution process and significant effects of the adaptive parameter adjustment strategy in this embodiment. Through simulated annealing, the algorithm achieves precise, phased control of parameter configuration: In the initial stage, the algorithm assigns a dominant exploration weight of 0.85, coupled with a large parameter update step size of 0.12, ensuring extensive exploration of the solution space; as the iteration enters the middle stage, the exploration and development weights tend to balance, and the update step size is appropriately reduced; in the later stage, the development weight increases to 0.77, and the update step size decreases to 0.03, focusing on refined local optimization; this dynamic adjustment strategy increases the convergence speed by 3.7 times in the later stage compared to the initial stage, with an average speedup of 2.8 times; this mechanism effectively balances the contradiction between global exploration and local development, allowing the algorithm to maintain sufficient randomness in the early search to avoid premature convergence, while efficiently and accurately locating the optimal solution in the later stage, greatly improving the overall performance and adaptability of the algorithm.

[0105] Specifically, this embodiment integrates the advantages of quantum particle swarm optimization (QPSO) and genetic algorithms to construct a multi-level parallel search strategy. The quantum encoding strategy represents parameters as probability distributions rather than discrete values, enabling a single particle to simultaneously explore multiple solution space locations, expanding the search range by 3.5 times. The introduced adaptive crossover and mutation mechanism dynamically adjusts operator parameters based on population diversity, avoiding early convergence risks and improving global search capability by 42.7%. Pareto dominance ranking and crowding calculation achieve precise quantitative evaluation of multi-objective trade-offs, resolving the performance conflict problem between different frequency bands in traditional methods. Particularly in adaptive weight adjustment, the simulated annealing mechanism assigns higher weights to exploration in the early stages and gradually shifts towards fine-tuning in the later stages, improving convergence speed by 63.8%. Experimental verification shows that, under the same computational resources, this hybrid framework reduces the number of iterations by 58.2% compared to a single evolutionary algorithm, and improves the overall performance index of the final solution by 27.5%, providing an efficient and reliable computational method for simultaneous optimization of multi-band performance of metasurface antennas.

[0106] The electromagnetic characteristics of the extracted metasurface antenna parameter combination are then iteratively adjusted based on the gain-impedance-bandwidth matching degree until a threshold is met, yielding electromagnetic simulation verification results. A parameter optimization closed loop is constructed using a reinforcement learning mechanism, and the electromagnetic simulation verification results are fed back to a deep Q-neural network model for parameter updates, resulting in the optimal metasurface antenna parameter combination, including:

[0107] A real-time electromagnetic field distribution feature extraction mechanism is constructed. Electromagnetic simulation is performed on the combination of metasurface antenna parameters generated by the hybrid optimization algorithm to obtain electromagnetic field distribution maps and S-parameter curves in real time. Electromagnetic features are extracted through a pre-trained convolutional neural network model. An iterative adjustment strategy for the resonant cavity height parameter is constructed to adjust the resonant cavity height according to the gain and impedance bandwidth matching degree.

[0108] In one specific embodiment, the real-time electromagnetic field distribution feature extraction and resonant cavity height iterative adjustment include:

[0109] For each set of antenna parameters generated by the hybrid optimization algorithm, a three-dimensional model is constructed using electromagnetic simulation software, and its electromagnetic response is calculated. A pre-trained convolutional neural network is used to extract the spatial structural features of the electromagnetic field distribution map in real time, including the standing wave ratio, radiation pattern, and electric field distribution gradient. An adaptive grid sampling strategy is designed to perform non-uniform sampling of the resonant cavity height parameter space, using a finer grid in electromagnetically sensitive areas. A gain-bandwidth tradeoff model is established, defining the matching degree index between gain and impedance bandwidth. Combining gradient information and pattern recognition technology, the optimal resonant cavity height adjustment direction and step size are automatically identified. A Gaussian process regression model is introduced to learn the relationship between resonant cavity height and antenna performance online, predicting the optimal parameter range and dynamically narrowing the search space. The posterior probability distribution of the resonant cavity height is continuously updated through a Bayesian optimization framework to balance exploration and utilization until the matching threshold requirements of gain and impedance bandwidth are met.

[0110] A parameter optimization closed loop based on deep reinforcement learning is constructed. The comprehensive evaluation function of antenna performance is used as the reward signal to establish a state-action-reward mapping relationship. The optimization strategy of metasurface antenna parameters is learned through deep Q neural network. The electromagnetic simulation verification results are fed back to the model weight update process to perform adaptive closed-loop optimization between metasurface antenna parameters and electromagnetic characteristics.

[0111] In one specific embodiment, the parameter optimization closed loop based on deep reinforcement learning includes:

[0112] The antenna parameter optimization problem is modeled as a Markov decision process, with the state space consisting of the current antenna parameter combination and extracted electromagnetic feature vectors, and the action space defined by the adjustment amount of the antenna parameters. A hierarchical reward function is designed, including immediate and delayed rewards. The immediate reward is based on the performance change before and after a single parameter adjustment, while the delayed reward is based on the degree to which the final antenna performance reaches the design target. A deep Q-learning algorithm with a dual-network architecture is adopted, including an evaluation network and a target network. The evaluation network is used for action selection, and the target network is used to stabilize the learning process. A priority experience replay mechanism is introduced, assigning higher sampling probabilities to adjustment experiences that contain significant performance improvements. An exploration noise annealing strategy is designed to gradually reduce the randomness of parameter exploration as the learning process progresses. An experience knowledge base is constructed to store the parameter-performance mapping relationship in the historical optimization process, which is used to assist model training and reduce unnecessary simulation calculations. By jointly optimizing the policy gradient and the value function, the model can adaptively adjust the parameter optimization direction while maintaining stable convergence. After multiple iterations of training, the optimal antenna parameter adjustment strategy is formed, realizing automatic optimization of electromagnetic performance.

[0113] Specifically, this embodiment achieves precise iterative adjustment of the height parameter of the metasurface antenna resonator by constructing a parameter optimization closed-loop mechanism based on deep reinforcement learning. This embodiment utilizes a deep Q-neural network combined with an experience replay mechanism to automatically evaluate the current electromagnetic field state and predict the optimal parameter adjustment strategy. The reward function precisely quantifies the improvement in gain and impedance bandwidth matching, establishing a complete closed loop of simulation verification, feature extraction, parameter adjustment, and performance evaluation. This embodiment significantly reduces the number of electromagnetic simulation calculations, accelerates the optimization convergence process, and achieves knowledge transfer and experience accumulation through the cumulative learning of the memory bank, forming a parameter optimization system with adaptive learning capabilities.

[0114] Please see Figure 2 The present invention also provides a metasurface antenna parameter optimization system based on a convolutional neural network, the system comprising:

[0115] The sample data acquisition module is used to collect raw sample data of the metasurface antenna and preprocess it to obtain metasurface sample data.

[0116] The mapping model building module is used to construct a two-stream convolutional neural network model. Through the two-stream convolutional neural network model, spatial domain features and frequency domain features are extracted from the metasurface sample data. Based on the spatial domain features and frequency domain features, a comprehensive electromagnetic feature set is constructed, and a nonlinear mapping relationship model between electromagnetic features and resonant cavity height is established.

[0117] The antenna parameter optimization module is used to construct a comprehensive evaluation function for antenna performance. The nonlinear mapping relationship model and the comprehensive evaluation function for antenna performance are input into a hybrid optimization algorithm. A parameter candidate set is generated through a genetic algorithm, and the parameter candidate set is locally optimized using a particle swarm optimization algorithm to obtain the combination of metasurface antenna parameters.

[0118] The simulation verification feedback module is used to extract the electromagnetic characteristics of the metasurface antenna parameter combination. Based on the matching degree of gain and impedance bandwidth, the resonant cavity height parameter is iteratively adjusted until a threshold is met to obtain the electromagnetic simulation verification result. A parameter optimization closed loop is constructed through a reinforcement learning mechanism, and the electromagnetic simulation verification result is fed back to the deep Q neural network model for parameter update to obtain the optimal metasurface antenna parameter combination.

[0119] Specifically, this embodiment of a metasurface antenna parameter optimization system based on convolutional neural networks integrates a sample data acquisition module, a mapping model establishment module, an antenna parameter optimization module, and a simulation verification feedback module, achieving automation and efficiency throughout the entire metasurface antenna parameter optimization process. The system supports parameter optimization tasks for different antenna structure types, provides intuitive visualization analysis functions to show the correlation between electromagnetic characteristics and parameters, and has a built-in knowledge base for accumulating and transferring design experience. In practical engineering applications, the system significantly shortens the antenna design cycle, reduces design costs and professional knowledge barriers, improves the quality and reliability of design schemes, and has good scalability and deployment convenience, providing powerful tool support for the rapid application of metasurface antenna technology in 5G / 6G communication systems, aerospace, and smart IoT fields.

[0120] The present invention also discloses an electronic device, comprising: at least one processor, at least one memory communication interface and bus; wherein the processor, memory and communication interface communicate with each other through the bus; the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to implement a metasurface antenna parameter optimization method based on a convolutional neural network.

[0121] This invention also discloses a computer-readable storage medium that stores computer instructions that cause the computer to implement all or part of the steps of the metasurface antenna parameter optimization method based on a convolutional neural network described in this embodiment of the invention. The storage medium includes various media capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for optimizing parameters of a metasurface antenna based on a convolutional neural network, characterized in that, The method comprises the following steps: Collecting and preprocessing original sample data of the metasurface antenna to obtain metasurface sample data; A dual-stream convolutional neural network model is constructed to extract spatial domain features and frequency domain features from the metasurface sample data, construct a comprehensive electromagnetic feature set based on the spatial domain features and the frequency domain features, and establish a nonlinear mapping relationship model between electromagnetic features and resonant cavity height; The construction of the dual-stream convolutional neural network model, the extraction of spatial domain features and frequency domain features from the metasurface sample data by the dual-stream convolutional neural network model, the construction of a comprehensive electromagnetic feature set based on the spatial domain features and the frequency domain features, and the establishment of a nonlinear mapping relationship model between electromagnetic features and resonant cavity height comprise: A dual-stream convolutional neural network is constructed, which includes a spatial domain feature extraction branch and a frequency domain feature extraction branch. The spatial domain feature extraction branch uses three-dimensional convolution to process electromagnetic field distribution maps, and the frequency domain feature extraction branch uses one-dimensional convolution to process S parameter curves. The features of the two branches are merged through a feature fusion layer to construct a comprehensive electromagnetic feature set; A nonlinear mapping model is trained using the comprehensive electromagnetic feature set and corresponding resonant cavity height data. The nonlinear mapping model adopts a multilayer perceptron structure; An antenna performance comprehensive evaluation function is constructed. The nonlinear mapping relationship model and the antenna performance comprehensive evaluation function are input into a hybrid optimization algorithm. A parameter candidate set is generated by a genetic algorithm, and the parameter candidate set is locally optimized by a particle swarm algorithm to obtain a metasurface antenna parameter combination; The electromagnetic features of the metasurface antenna parameter combination are extracted. The resonant cavity height parameter is iteratively adjusted based on the matching degree of gain and impedance bandwidth until a threshold is met, and an electromagnetic simulation verification result is obtained. A parameter optimization closed loop is constructed through a reinforcement learning mechanism. The electromagnetic simulation verification result is fed back to a deep Q neural network model for parameter updating to obtain an optimal metasurface antenna parameter combination.

2. The metasurface antenna parameter optimization method based on convolutional neural network according to claim 1, wherein, The collection and preprocessing of original sample data of the metasurface antenna to obtain metasurface sample data comprise: Different structure parameters and resonant cavity heights of the metasurface antenna are simulated by electromagnetic simulation to obtain corresponding electromagnetic field distribution maps and S parameter curves, and original sample data of the metasurface antenna are obtained; The original sample data are standardized and enhanced to eliminate dimensional differences. Data denoising, missing value processing, and image enhancement are performed to improve data quality, and a metasurface sample data set for training is constructed. The metasurface sample data include antenna structure parameters, electromagnetic field distribution maps, and S parameter curves.

3. The metasurface antenna parameter optimization method based on convolutional neural network according to claim 1, wherein, The construction of the dual-stream convolutional neural network includes a spatial domain feature extraction branch and a frequency domain feature extraction branch. The spatial domain feature extraction branch adopts a three-dimensional convolution network with a residual structure, which is composed of three convolution units, each of which includes two 3*3*3 convolution layers, a batch normalization layer and a PReLU activation function, and residual connection is set, and the number of convolution kernels increases from 32 at the beginning to 128 layer by layer, the frequency domain feature extraction branch adopts an attention mechanism enhanced one-dimensional convolution network, including a time attention module and a frequency attention module, the frequency domain feature extraction branch uses multi-scale convolution kernels to process S parameter curves in parallel, and captures the antenna response characteristics in different frequency domain ranges; the feature fusion layer adopts an adaptive weighting mechanism, integrates the features extracted by the spatial domain feature extraction branch and the frequency domain feature extraction branch, and dynamically adjusts the weights of different features according to the importance of a specific frequency band.

4. The metasurface antenna parameter optimization method based on convolutional neural network of claim 1, wherein, The nonlinear mapping relationship model and the antenna performance comprehensive evaluation function are input into a hybrid optimization algorithm, a parameter candidate set is generated by a genetic algorithm, and the parameter candidate set is locally optimized by a particle swarm algorithm to obtain a metasurface antenna parameter combination, including: A multi-objective antenna performance comprehensive evaluation function is constructed, a frequency band adaptive weight coefficient is introduced, and a differential evaluation standard is set for the bandwidth, gain, directivity and impedance matching parameters of different frequency bands for weighted calculation to obtain an antenna performance comprehensive evaluation function; A hybrid optimization algorithm framework is established, the nonlinear mapping relationship model and the antenna performance comprehensive evaluation function set are integrated, a metasurface antenna parameter candidate set is generated by a genetic algorithm, the parameter space exploration ability of the metasurface antenna parameter candidate set is enhanced by an adaptive mutation strategy, a plurality of candidate parameters are obtained, and the candidate parameters are locally optimized based on a quantum particle swarm algorithm to obtain a metasurface antenna parameter combination.

5. The metasurface antenna parameter optimization method based on convolutional neural network according to claim 4, wherein, The construction of the antenna performance comprehensive evaluation function includes: For the working frequency band characteristics of the metasurface antenna, the target frequency range is divided into a plurality of sub-frequency bands, different priorities are set for each frequency band, the normalized bandwidth index, gain index, directivity index and impedance matching index are defined for each frequency band, the coupling effect between frequency bands is identified by frequency band correlation analysis, a frequency band performance balance model is established, a hierarchical evaluation structure is constructed for multi-frequency band application scenarios, the redundancy between evaluation indexes is reduced by principal component analysis, key performance characteristics are extracted, a fuzzy comprehensive evaluation method is used to integrate the evaluation results of multiple frequency bands and multiple indexes to obtain a single antenna performance comprehensive evaluation function; The establishment of the hybrid optimization algorithm framework includes: After initializing the population, an improved genetic algorithm with adaptive crossover and mutation strategies is used to generate a candidate set of antenna parameters, the crossover probability and mutation probability are dynamically adjusted according to the population diversity and fitness distribution, the multi-objective optimization problem is processed through a Pareto dominance sorting mechanism, the non-dominated solution set is retained, for each candidate solution in the non-dominated solution set, the optimal resonant cavity height is predicted using a nonlinear mapping relationship model; the predicted optimal resonant cavity height is combined with the metasurface antenna parameter to obtain a complete solution, and a quantum particle swarm optimization algorithm is used for local optimization, wherein the quantum rotation angle and the contraction expansion factor are adaptively adjusted based on the quality of the complete solution; a dynamic balance mechanism for solution space exploration and development is set, and the proportion of global search and local refinement is adjusted according to the optimization process.

6. The metasurface antenna parameter optimization method based on convolutional neural network according to claim 5, wherein, The electromagnetic characteristics of the extracted metasurface antenna parameter combination are iteratively adjusted based on the matching degree of the gain and impedance bandwidth until a threshold is met, and the electromagnetic simulation verification result is obtained, a parameter optimization closed loop is constructed through a reinforcement learning mechanism, the electromagnetic simulation verification result is fed back to a deep Q neural network model for parameter updating, and the optimal metasurface antenna parameter combination is obtained, including: An real-time electromagnetic field distribution feature extraction mechanism is constructed, the metasurface antenna parameter combination generated by the hybrid optimization algorithm is subjected to electromagnetic simulation, the electromagnetic field distribution map and the S parameter curve are obtained in real time, the electromagnetic characteristics are extracted through a pre-trained convolutional neural network model, an iterative adjustment strategy for the resonant cavity height parameter is constructed, and the resonant cavity height is adjusted according to the matching degree of the gain and the impedance bandwidth; A parameter optimization closed loop based on deep reinforcement learning is constructed, the antenna performance comprehensive evaluation function is used as a reward signal, a state-action-reward mapping relationship is established, a deep Q neural network is used to learn the metasurface antenna parameter optimization strategy, the electromagnetic simulation verification result is fed back to the model weight updating process, and adaptive closed loop optimization is performed between the metasurface antenna parameter and the electromagnetic characteristics.

7. A convolutional neural network based metasurface antenna parameter optimization system configured to perform a convolutional neural network based metasurface antenna parameter optimization method according to any one of claims 1-6, wherein The system comprises: A sample data acquisition module is configured to collect original sample data of a metasurface antenna and perform preprocessing to obtain metasurface sample data; A mapping model establishment module is configured to construct a dual-stream convolutional neural network model, extract spatial domain features and frequency domain features from the metasurface sample data through the dual-stream convolutional neural network model, construct a comprehensive electromagnetic feature set based on the spatial domain features and the frequency domain features, and establish a nonlinear mapping relationship model between the electromagnetic features and the resonant cavity height; An antenna parameter optimization module is configured to construct an antenna performance comprehensive evaluation function, input the nonlinear mapping relationship model and the antenna performance comprehensive evaluation function into a hybrid optimization algorithm, generate a parameter candidate set through a genetic algorithm, and perform local optimization on the parameter candidate set through a particle swarm optimization algorithm to obtain a metasurface antenna parameter combination; An simulation verification feedback module is configured to extract electromagnetic characteristics of the metasurface antenna parameter combination, iteratively adjust the resonant cavity height parameter based on the matching degree of the gain and the impedance bandwidth until a threshold is met, obtain an electromagnetic simulation verification result, construct a parameter optimization closed loop through a reinforcement learning mechanism, feed back the electromagnetic simulation verification result to a deep Q neural network model for parameter updating, and obtain an optimal metasurface antenna parameter combination.

8. An electronic device, comprising: The computer readable storage medium stores computer instructions, and the computer instructions cause the computer to implement the method in any one of claims 1-6. The computer readable storage medium stores computer instructions, and the computer instructions cause the computer to implement the method in any one of claims 1-6. ​ 9. A computer-readable storage medium, characterized in that, ​

Images