A method, apparatus, device and medium for design optimization of a MIMO antenna

Abstract

The application provides a MIMO antenna design optimization method, device, equipment and medium, and belongs to the antenna design field. The method comprises the following steps: obtaining historical design parameters and corresponding performance indicators of a MIMO antenna; preprocessing the historical design parameters and the performance indicators, constructing a mapping relationship between input data and output data, and training a forward prediction model and a generative model based on the mapping relationship; receiving a target performance indicator input, generating a candidate design parameter set by using the trained generative model, taking the candidate design parameter set as an initial population, combining the forward prediction model as a surrogate model, adopting a multi-objective evolutionary algorithm for iterative optimization, and outputting a Pareto optimal solution set as a target design parameter set. In this way, the application greatly speeds up the performance evaluation link with the help of the surrogate model, provides diversified non-inferior solution selection through Pareto frontier analysis, and can flexibly trade off different design targets according to actual requirements.

Description

Technical Field

[0001] This invention belongs to the field of antenna design, specifically relating to a design optimization method, apparatus, device, and medium for MIMO antennas. Background Technology

[0002] With the rapid development of fifth-generation (5G) and future sixth-generation (6G) mobile communication systems, the requirements for data transmission rate, spectral efficiency, and connection reliability are increasing. Multiple-input multiple-output (MIMO) antenna technology, by deploying multiple antenna elements at both the transmitting and receiving ends and utilizing spatial multiplexing and diversity gain, has become one of the key technologies for improving the performance of wireless communication systems.

[0003] The performance of MIMO antennas (such as gain, bandwidth, radiation pattern, isolation, efficiency, etc.) has a complex and highly nonlinear relationship with their physical structure (such as element shape, size, spacing, feed network, substrate material, etc.). The design process of traditional MIMO antennas largely depends on the experience of design engineers and repeated electromagnetic (EM) simulation iterations.

[0004] However, this method has significant drawbacks: it has a long calculation cycle and high cost. It mainly relies on manual experience and trial and error for parameter adjustment, and verification is carried out through time-consuming full-wave electromagnetic simulation, resulting in a lengthy R&D cycle and high costs for human and computing resources. Moreover, MIMO antenna design involves many interrelated parameters, forming a huge design space. Traditional optimization methods (such as parameter scanning, gradient descent, etc.) are prone to getting trapped in local optima and find it difficult to find the global optimum under the conditions of satisfying multiple performance indicators (such as high gain, wide bandwidth, high isolation, small size, etc.) and physical constraints, which further leads to low optimization efficiency. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method, apparatus, device, and medium for designing and optimizing MIMO antennas.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A design optimization method for MIMO antennas, the method comprising:

[0008] Obtain the historical design parameters and corresponding performance indicators of the MIMO antenna. The design parameters include geometric parameters, material parameters, and feed network parameters. The performance indicators include S-parameters, radiation characteristics, MIMO characteristics, and bandwidth.

[0009] Historical design parameters are encoded to generate structured input data, performance indicators are normalized to generate structured output data, a mapping relationship between input and output data is constructed, and a forward prediction model and a generative model are trained based on the mapping relationship; the forward prediction model is used to predict performance indicators through design parameters; the generative model is used to predict design parameters through performance indicators.

[0010] Different target performance indicators are received as inputs, and a candidate design parameter set is generated using a trained generative model. The candidate design parameter set is used as the initial population, and the corresponding predicted performance indicator set is predicted by a forward prediction model. The Pareto dominance relationship of the predicted performance indicator set is determined, and multi-objective optimization is performed on the initial population based on the Pareto dominance relationship to obtain the target design parameter set.

[0011] Optionally, the step of combining the forward prediction model as a surrogate model and using a multi-objective evolutionary algorithm for iterative optimization includes:

[0012] The predicted performance index set corresponding to the candidate design parameter set is obtained by using a surrogate model, and the predicted performance index set is converted into fitness values ​​based on the Pareto dominance relationship.

[0013] Based on the fitness value, the candidate design parameter set is iteratively optimized using a multi-objective evolutionary algorithm.

[0014] Optionally, the process of encoding historical design parameters to generate structured input data and normalizing performance indicators to generate structured output data includes:

[0015] Vector encoding, image / voxel matrix encoding, or graph structure encoding are used for geometric parameters;

[0016] Material parameters can be input directly using numerical values ​​or encoded using unique thermal methods.

[0017] Frequency-related performance indicators are sampled as fixed-length vectors.

[0018] Optionally, the graph structure encoding includes:

[0019] The antenna elements in the MIMO array are treated as graph nodes. The node features include the element's own parameters, and the edge features include the estimated values ​​of the element spacing or coupling strength.

[0020] Optionally, training the forward prediction model and the generative model based on the mapping relationship includes:

[0021] By training a CNN or GNN, a nonlinear mapping from design parameters to performance metrics is learned, resulting in a forward prediction model; by training a VAE or GAN, an inverse mapping from performance metrics to design parameters is learned, resulting in a generative model; wherein, the CNN is used to process the geometric structure of image / voxel encoding and extract spatial features; the GNN is used to process the array parameters of graph structure encoding and explicitly model the coupling effect between units.

[0022] Optionally, generating a candidate design parameter set using the trained generative model includes:

[0023] Input the target performance index into the conditional generative adversarial network (GAN) to directly generate candidate design parameters;

[0024] Alternatively, sampling optimization can be performed using the latent space of a variational autoencoder (VAE):

[0025] The target performance is mapped to the target region of the potential space, potential vectors are sampled in the target region, and candidate design parameters are generated based on the potential vectors through the VAE decoder.

[0026] A design optimization device for a MIMO antenna, characterized in that the device comprises:

[0027] The acquisition module is used to acquire the historical design parameters and corresponding performance indicators of the MIMO antenna. The design parameters include geometric parameters, material parameters, and feed network parameters; the performance indicators include S-parameters, radiation characteristics, MIMO characteristics, and bandwidth.

[0028] A construction module is used to encode historical design parameters to generate structured input data, normalize performance indicators to generate structured output data, construct a mapping relationship between input data and output data, and train a forward prediction model and a generative model based on the mapping relationship; the forward prediction model is used to predict performance indicators through design parameters; the generative model is used to predict design parameters through performance indicators.

[0029] The optimization module is used to receive different target performance indicators as inputs, generate a set of candidate design parameters using the trained generative model, use the set of candidate design parameters as the initial population, and predict the corresponding set of predicted performance indicators through the forward prediction model, determine the Pareto dominance relationship of the set of predicted performance indicators, and perform multi-objective optimization on the initial population based on the Pareto dominance relationship to obtain the set of target design parameters.

[0030] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned design optimization method for a MIMO antenna.

[0031] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned design optimization method for a MIMO antenna.

[0032] The design optimization method for MIMO antennas provided by this invention has the following beneficial effects:

[0033] This invention first utilizes historical antenna design data to construct a bidirectional prediction model. This data preprocessing and model training mechanism effectively overcomes the limitations of traditional trial-and-error methods, eliminating the need for extensive repetitive simulations in the optimization process. Then, by combining the high-quality initial solutions output by the generated model with a multi-objective evolutionary algorithm, the system can efficiently search for the globally optimal solution set within the design space, while automatically balancing conflicts among multiple dimensions such as radiation characteristics and bandwidth. Compared to traditional methods that are prone to getting trapped in local optima and have high computational costs, this invention significantly accelerates the performance evaluation process using a surrogate model and provides diverse non-dominated solution choices through Pareto front analysis. This allows for flexible balancing of different design objectives according to actual needs, improving design efficiency. This method is particularly suitable for MIMO antenna designs that require consideration of complex electromagnetic performance, and its modular architecture also shows potential for expansion into other electronic device optimization scenarios. Attached Figure Description

[0034] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. 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.

[0035] Figure 1 This is a flowchart illustrating a design optimization method for a MIMO antenna according to an exemplary embodiment of the present invention.

[0036] Figure 2 This is a schematic diagram of MIMO antenna parameter encoding according to an exemplary embodiment of the present invention.

[0037] Figure 3 This is a structural diagram of a deep learning model for MIMO antenna design provided by the present invention according to an exemplary embodiment.

[0038] Figure 4 This is a flowchart of a MIMO antenna reverse engineering engine provided by the present invention according to an exemplary embodiment.

[0039] Figure 5 This is a block diagram of a design optimization device for a MIMO antenna according to an exemplary embodiment of the present invention. Detailed Implementation

[0040] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0041] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0042] First, this invention provides a design optimization method for MIMO antennas, specifically as follows: Figure 1 As shown, it includes the following steps:

[0043] S101. Obtain the historical design parameters and corresponding performance indicators of the MIMO antenna.

[0044] The design parameters include geometric parameters, material parameters, and feed network parameters; the performance indicators include S-parameters, radiation characteristics, MIMO characteristics, and bandwidth.

[0045] In this step, the original MIMO antenna design parameters mainly include the following parameters:

[0046] Geometric parameters: length, width, thickness, slot size, and feed point location of antenna elements (such as patches and dipoles); array parameters such as element spacing, arrangement (linear, planar), and number of elements.

[0047] Material parameters: substrate dielectric constant, loss tangent, thickness; conductor material conductivity.

[0048] Feed network parameters: impedance matching network component values, feeder characteristic impedance, power divider / coupler structural parameters.

[0049] Environmental parameters: such as the properties of the packaging materials around the antenna.

[0050] These parameters can come from historical design databases, literature, or be generated using parametric modeling software (such as AutoCAD, SolidWorks, and scripts).

[0051] Performance data obtained through simulation of the above design parameter combinations using electromagnetic (EM) simulation software (such as ANSYS HFSS, CST Studio Suite), or through actual measurements. Performance indicators include, but are not limited to:

[0052] S-parameters: return loss (S11, S22, ...), isolation (S12, S13, S21, ...), transmission coefficient.

[0053] Radiation characteristics: gain, directivity coefficient, efficiency (radiation efficiency, overall efficiency), half-power beamwidth, front-to-back ratio, cross-polarization ratio, sidelobe level, radiation intensity in a specific direction.

[0054] MIMO characteristics: envelope correlation coefficient (ECC), channel capacity loss (CCL), diversity gain, and average effective gain (MEG).

[0055] Bandwidth: Impedance bandwidth (e.g., frequency range where S11 < -10dB), gain bandwidth, efficiency bandwidth.

[0056] S102. Encode the historical design parameters to generate structured input data, normalize the performance indicators to generate structured output data, construct the mapping relationship between the input data and the output data, and train the forward prediction model and the generative model based on the mapping relationship.

[0057] The forward prediction model is used to predict performance metrics based on design parameters; the generative model is used to predict design parameters based on performance metrics. For example, the forward prediction model can be obtained by training a CNN or GNN to learn a nonlinear mapping from design parameters to performance metrics; the generative model can be obtained by training a VAE or GAN to learn an inverse mapping from performance metrics to design parameters. The CNN is used to process the geometric structure of image / voxel encoding and extract spatial features; the GNN is used to process the array parameters of graph structure encoding and explicitly model the coupling effects between units.

[0058] This step begins with data preprocessing. This step is the starting point of the entire system, responsible for preparing and transforming the raw data to make it suitable for training deep learning models.

[0059] The preprocessing of the original MIMO antenna design parameters mainly includes the following steps: vector encoding, image / voxel matrix encoding, or graph structure encoding for geometric parameters; direct numerical input or one-hot encoding for material parameters; and sampling of frequency-related performance indicators as fixed-length vectors.

[0060] For parameter encoding, such as Figure 2 As shown, it mainly includes:

[0061] Vector encoding: For antennas with relatively simple structures or a small number of parameters, all geometric dimensions and material parameters are directly arranged into a one-dimensional numerical vector.

[0062] Image / voxel encoding: For complex geometries, their two-dimensional projection or three-dimensional structure can be discretized into pixel / voxel matrices, which can be used as input to CNN.

[0063] Graph structure encoding (for arrays): Each antenna element in the MIMO antenna array is regarded as a graph node. The node features can include the parameters of the element itself (such as size and feed); the edge features can include the spacing between elements (physical distance or relative position between elements) or the estimated value of coupling strength (the weight of the edge can be initially estimated as the coupling strength, and this encoding method is particularly suitable for GNN).

[0064] For example, the patch size and feed point location can be discretized into a 128×128 pixel matrix; the cell spacing and arrangement can be mapped into a weighted edge attribute graph.

[0065] For material parameter encoding, the main methods are: continuous parameters (such as dielectric constant) are directly represented by numerical values; discrete parameters (such as those selected from a set of pre-selected materials) can be represented by one-hot encoding or embedding vector.

[0066] For performance index coding, the main categories are: scalar performance (such as gain and isolation at a specific frequency) can be used directly; frequency-dependent performance (such as S-parameter curves and gain versus frequency curves) can be sampled as fixed-length vectors (sampled at a specific frequency point) or represented by function approximation coefficients.

[0067] The next step is data cleaning, which involves identifying and processing data points that failed in the simulation, are missing, or are obviously abnormal (e.g., based on physical principles or statistical methods such as IQR). Processing methods may include deletion, interpolation (mean, median, model prediction filling), etc.

[0068] Then, data normalization / standardization is performed to scale all numerical input parameters and output performance metrics to eliminate the influence of units and accelerate model convergence. Common methods include:

[0069] Min-Max Scaling: Scaling the data to the range [0,1] or [-1,1].

[0070] Z-score standardization transforms data into a distribution with a mean of 0 and a standard deviation of 1.

[0071] In addition, data augmentation can be performed. For datasets with small sizes, existing design parameters can be slightly perturbed (e.g., randomly adjusting dimensional parameters by ±5% within a reasonable range), and the performance can be re-simulated to expand the dataset. Alternatively, a trained generative model (such as a VAE) can be used to generate similar but not identical design parameters and predicted performance, which can then be added to the dataset after simulation verification.

[0072] The final output is a structured mapping database: a database containing multiple pairs of (X, Y) samples, where X is an encoded and normalized / standardized vector / matrix / graph of antenna design parameters, and Y is the corresponding encoded and normalized / standardized performance metric vector. This database will serve as the foundation for subsequent deep learning model training.

[0073] In one embodiment, the geometric feature vector of the image matrix and the unit coupling relationship matrix of the graph structure are extracted respectively; the geometric feature vector and the unit coupling relationship matrix are concatenated to generate a parameter matrix; based on the parameter matrix, multiple sets of candidate design parameters are generated by a deep learning generation model under multiple preset target performance conditions, and the performance indicators of each set of candidate design parameters corresponding to multiple preset target performance are predicted by a pre-trained surrogate model.

[0074] For candidate design parameters, the image matrix is ​​first input into a CNN model for feature extraction to obtain geometric feature vectors; the image structure is then input into a GNN model for feature extraction to obtain a unit coupling matrix. Next, using a preset target performance as a conditional vector, and the geometric feature vectors and coupling matrix as joint inputs, the deep learning generative model generates initial design parameters and corresponding reconstruction target performance metrics. Gradient descent is then performed to minimize the difference between the reconstruction target performance metrics and the preset target performance, outputting the initial design parameters. Finally, the initial design parameters are filtered based on preset physical conditions to obtain candidate design parameters. For example, the deep learning generative model can include an encoder and a decoder. The encoder generates the initial design parameters, and the decoder generates the corresponding reconstruction target performance metrics. The difference between the reconstruction target performance metrics and the preset target performance is then obtained. The deep learning generative model is trained with the goal of minimizing this difference. If the difference is less than a preset difference, training is considered complete, and the initial design parameters are output.

[0075] By utilizing data obtained from the data preprocessing module, a series of deep learning models are trained that can work together to accurately capture the nonlinear mapping relationship between antenna design parameters and their complex electromagnetic properties, and support efficient design space exploration and generation, such as... Figure 3 As shown.

[0076] The structured mapping relation database (X, Y) obtained from the data preprocessing in the above steps is input into the pre-built forward prediction model.

[0077] The construction and training of a forward prediction model mainly includes the following steps:

[0078] Objective: To learn the mapping function f: X → Y from design parameters X to performance metrics Y.

[0079] CNN structural feature extractor (for image / voxel encoding input).

[0080] The architecture consists of multiple convolutional layers (extracting local features), activation functions (such as ReLU, introducing non-linearity), pooling layers (reducing dimensionality and parameters), and fully connected layers (integrating features and outputting prediction performance). The size, number, and stride of the convolutional kernels, as well as the pooling window size, are set according to the input image / voxel resolution and the complexity of the antenna structure.

[0081] Training: The input is an image / voxel representation of the antenna structure, and the output is the predicted performance metric. The optimization objective is to minimize the mean square error (MSE) or other suitable loss function between the predicted performance and the actual simulation / measurement performance.

[0082] GNN element relation modeler (for graph encoding input, especially for arrays).

[0083] The architecture includes graph convolutional layers or graph attention layers (aggregating neighbor node information to update node representations, thereby capturing inter-cell coupling), node update functions, and a graph readout layer (which aggregates node representations into graph-level representations for predicting overall array performance).

[0084] Training: The input is a graph representation of the antenna, and the output is the predicted array performance metrics (such as isolation and ECC). The optimization objective is also to minimize the difference between the prediction and the actual performance.

[0085] Hybrid models: For some complex situations, the features extracted by CNN / GNN can be concatenated with the parameters encoded by other vectors, and then passed through several fully connected layers for final performance prediction.

[0086] Then the generative model is built and trained.

[0087] Objective: To learn a mapping function g: Y_target → X_generated from the target performance metric Y_target (or random noise / conditions) to the design parameters X_generated that produce that performance, or to learn a latent representation of the parameter space.

[0088] Antenna design parameter space generator based on variational autoencoder (VAE):

[0089] The structure consists of an encoder and a decoder. The encoder compresses the input design parameters X into a low-dimensional latent vector z (usually following a Gaussian distribution); the decoder reconstructs the design parameters X' from the latent vector z.

[0090] Training: The optimization objective consists of two parts: reconstruction loss (e.g., MSE between X and X') and KL divergence between the latent space distribution and the prior distribution (e.g., the standard normal distribution N(0,I)). This makes the latent space smooth and continuous.

[0091] Generative networks based on Generative Adversarial Networks (GANs) are used for more powerful generation capabilities:

[0092] The architecture consists of a generator and a discriminator. The generator attempts to produce realistic design parameters from random noise (or conditional information, such as target performance); the discriminator attempts to distinguish between real design parameters and those generated by the generator.

[0093] Training: Through adversarial training, the generator learns to generate more realistic designs, and the discriminator learns to better distinguish between real and fake designs. In the case of a conditional GAN ​​(cGAN), the target performance is provided as additional input to the generator and discriminator.

[0094] A reinforcement learning (RL)-based antenna design optimization decision network is used to guide the optimization process:

[0095] Construction: Antenna design is viewed as a Markov Decision Process (MDP). The state of the agent (such as DQN, PPO) can be the current design parameters and their predicted performance, the action is to modify a design parameter (such as adding or removing a specific size), and the reward is designed based on the degree of performance improvement (such as the degree of proximity to the target, the amount of improvement in a multi-objective function).

[0096] Training: The agent learns a policy π (Action|State) to maximize cumulative rewards by interacting with the environment (which can be a fast AI prediction model).

[0097] S103. Receive different target performance indicators as inputs, generate a candidate design parameter set using the trained generative model, use the candidate design parameter set as the initial population, and predict the corresponding predicted performance indicator set through the forward prediction model, determine the Pareto dominance relationship of the predicted performance indicator set, and perform multi-objective optimization on the initial population based on the Pareto dominance relationship to obtain the target design parameter set.

[0098] Specifically, for the iterative optimization of the initial population, the forward prediction model can be used as a surrogate model to predict the predicted performance index set corresponding to the candidate design parameter set. Based on the Pareto dominance relationship, the predicted performance index set is converted into fitness values. Based on the fitness values, the candidate design parameter set is iteratively optimized using a multi-objective evolutionary algorithm.

[0099] A well-trained forward prediction model (CNN / GNN / hybrid model) can quickly and accurately predict its performance based on the input design parameters.

[0100] A well-trained generative model (VAE / GAN) can generate new antenna design parameters, or, given a target performance, generate design parameters that satisfy that performance. For example, the target performance index can be input into a conditional generative adversarial network (GAN) to directly generate candidate design parameters; alternatively, the latent space of a variational autoencoder (VAE) can be sampled and optimized: the target performance is mapped to a target region in the latent space, latent vectors are sampled in the target region, and candidate design parameters are generated based on the latent vectors through a VAE decoder.

[0101] A well-trained agent can guide the adjustment of design parameters.

[0102] In addition, a pre-trained generative model can be used to directly generate candidate antenna design schemes based on user-defined performance targets, such as... Figure 4 As shown.

[0103] User-specified target MIMO antenna performance metrics Y_target (e.g., operating frequency range, minimum gain, maximum S11, minimum isolation, target ECC, etc.).

[0104] A well-trained generative model (such as VAE, cGAN) or a network that performs inversely to the latent space.

[0105] Based on cGAN / cVAE: Y_target is used as a conditional input to the trained cGAN generator or cVAE decoder (in conjunction with latent vectors sampled from the conditional prior) to directly generate one or more sets of candidate design parameters X_candidate.

[0106] Then, VAE-based latent space optimization is performed: the target performance Y_target is mapped to the target region or point z_target in the latent space learned by VAE (an additional inverse mapping network from performance to latent space needs to be trained, or it can be achieved through optimization search).

[0107] Sample around z_target in the latent space, or sample the latent vector z from a standard normal distribution.

[0108] The sampled latent vectors are converted into candidate design parameters X_candidate by the VAE decoder.

[0109] Iterative refinement (e.g., using a diffusion model or combining it with a forward prediction model):

[0110] If a diffusion model is used, it starts with noise and, guided by Y_target, gradually denoises to generate X_candidate.

[0111] Alternatively, after generating an initial X_candidate, its performance Y_predicted is evaluated using a forward prediction model. The difference between Y_predicted and Y_target is calculated, and this difference (gradient) is used to guide the adjustment of X_candidate, iterating until the requirements are met.

[0112] Physical validity and constraint check: The generated X_candidate is checked to ensure that the parameter values ​​are within the physically feasible range (such as non-negative dimensions and optional materials) and meet the basic design rules, and invalid solutions are filtered out.

[0113] Finally, the output yields one or more sets of candidate MIMO antenna design parameters X_candidates. These parameters, predicted by the model, can meet or closely approximate the user-defined target performance Y_target. These parameters can be used as the initial population for a multi-objective optimizer or directly verified through simulation.

[0114] In one embodiment, a population can be initialized by constructing an initial population based on multiple sets of candidate design parameters; the population can be stratified based on non-dominated sorting according to the performance index of each individual, and Pareto front solutions can be preserved; a tournament selection mechanism can be used to select parent individuals, and offspring can be generated by simulating binary crossover and polynomial mutation, while constraining the repair operator during the mutation process; when the algorithm reaches a preset convergence threshold or the maximum number of iterations, a set of non-dominated solutions that achieves the optimal trade-off on multiple preset target performances and meets all constraints can be output as target design parameters.

[0115] This step is responsible for searching the design parameter space under various constraints to find a Pareto optimal solution that achieves the best trade-off among multiple conflicting performance objectives.

[0116] Initial population: It can be X_candidates generated from the reverse engineering engine, or a set of initial design parameters that are randomly generated or uniformly sampled based on design experience.

[0117] The trained forward prediction model (CNN / GNN) serves as a surrogate model for quickly evaluating the performance of candidate solutions.

[0118] Optimization objectives: A user-defined set of performance metrics and their optimization directions (maximization or minimization). For example: maximizing gain, minimizing S11 (within the operating frequency band), maximizing isolation, minimizing antenna size, and minimizing ECC.

[0119] Design constraints include physical size limitations, material selection range, frequency range, and manufacturing cost limitations.

[0120] Multi-objective evolutionary algorithms (MOEA) are employed, such as NSGA-II (Non-dominated Sorting Genetic Algorithm II), MOEA / D (Decomposition-based MOEA), and SPEA2 (Strength Pareto Evolutionary Algorithm 2).

[0121] Fitness assessment: For each individual in the population (i.e., a set of antenna design parameters), a trained forward prediction model (CNN / GNN) is used to quickly calculate its corresponding multiple performance index values.

[0122] Evolutionary operations include the following steps:

[0123] Selection: Based on criteria such as non-dominated ranking and crowding distance, select the best individuals to enter the next generation.

[0124] Crossover: Exchange parameters of the selected parent individual to generate a new child individual.

[0125] Mutation: Small random perturbations are made to certain parameters of offspring individuals to increase population diversity and explore new areas.

[0126] Constraint handling: For individuals that violate design constraints, a penalty function method (reducing their fitness) or a repair strategy is used, or they are removed directly from the population.

[0127] Adaptive sampling and proxy model update:

[0128] During the optimization process, a few representative or highly uncertain design options are selected from the current Pareto front periodically (e.g., every N generations, or when the surrogate model predicts high uncertainty).

[0129] These selected schemes are submitted to the simulation verification and feedback module for accurate EM simulation.

[0130] After obtaining the simulation results, they are added to the database of the data preprocessing module, and the deep learning model module is triggered to incrementally update or retrain the forward prediction model to improve the accuracy of the surrogate model.

[0131] Iteration: Repeat the steps of fitness evaluation, evolutionary operation, constraint handling, and adaptive sampling and surrogate model update until the termination condition is met (such as reaching the maximum number of iterations or Pareto front convergence and stability).

[0132] Finally, the output yields the Pareto Front: a set of MIMO antenna design parameter schemes that achieve different trade-offs among multiple objectives.

[0133] The above method first constructs a bidirectional prediction model using historical antenna design data. This data preprocessing and model training mechanism effectively overcomes the limitations of traditional trial-and-error methods, eliminating the need for extensive repetitive simulations in the optimization process. Then, by combining the high-quality initial solutions output by the generated model with a multi-objective evolutionary algorithm, the system can efficiently search for the globally optimal solution set within the design space, while automatically balancing conflicts among multiple dimensions such as radiation characteristics and bandwidth. Compared to traditional methods that are prone to getting trapped in local optima and have high computational costs, this invention significantly accelerates the performance evaluation process using a surrogate model and provides diverse non-dominated solution choices through Pareto front analysis, allowing for flexible trade-offs between different design objectives based on actual needs. This method is particularly suitable for MIMO antenna designs that require consideration of complex electromagnetic performance, and its modular architecture also demonstrates the potential for expansion into other electronic device optimization scenarios.

[0134] In one embodiment, simulation verification can also be performed based on the above steps.

[0135] This module connects the AI ​​system's optimized design with high-precision electromagnetic simulation software for final verification and forms a closed-loop feedback loop to continuously improve the AI ​​model.

[0136] One or more selected design schemes (antenna parameters) from the Pareto optimal solution set output by the multi-objective optimizer. The AI ​​model predicts the performance of these schemes.

[0137] Automated simulation interface:

[0138] The selected antenna design parameters are automatically converted into a format (such as script file or model file) that can be recognized by specific EM simulation software (such as ANSYS HFSS or CST StudioSuite).

[0139] The simulation task is set up automatically, including frequency range, port excitation, boundary conditions, mesh generation settings, etc.

[0140] Start and execute the full-wave electromagnetic simulation.

[0141] Result Analysis and Comparison:

[0142] After the simulation is completed, key performance indicators (S-parameters, radiation pattern data, ECC, etc.) are automatically extracted from the output file of the simulation software.

[0143] The actual performance obtained from the simulation is quantitatively compared with the predicted performance of the AI ​​model, and the error (such as absolute error and relative error) is calculated.

[0144] Database updates and model retraining are triggered as follows:

[0145] The simulation-verified design parameters and their corresponding real performance data pairs (X_verified, Y_simulated) are added back to the mapping database of the data preprocessing module.

[0146] Feedback Mechanism: If the simulation results differ significantly from the prediction results, or if a preset strategy is implemented (such as accumulating N new validation data points, fixed time intervals, or a decline in model performance on the validation set), the deep learning model module is triggered to retrain or fine-tune the relevant model (mainly the forward prediction model) using the updated database. This closed loop ensures that the model can continuously learn from new simulation data, constantly improving its accuracy and generalization ability.

[0147] The final MIMO antenna design scheme that has been verified: If the simulation results meet all design requirements, then this scheme is the final output.

[0148] Performance Comparison Report: Shows a comparison between AI prediction performance and simulated real-world performance.

[0149] The updated mapping database includes newly added simulation verification data points.

[0150] Model update command: Triggers the retraining of the deep learning model.

[0151] Through the collaborative work of the above modules, the AI-assisted MIMO antenna design optimization system of the present invention can realize a complete closed-loop process from data preparation, model training, reverse design, multi-objective optimization to simulation verification and model iteration.

[0152] Secondly, the present invention also provides a design optimization device for MIMO antennas, such as... Figure 5 As shown, it includes:

[0153] The acquisition module 201 is used to acquire the historical design parameters and corresponding performance indicators of the MIMO antenna. The design parameters include geometric parameters, material parameters, and feed network parameters; the performance indicators include S-parameters, radiation characteristics, MIMO characteristics, and bandwidth.

[0154] The construction module 202 is used to encode historical design parameters to generate structured input data, normalize performance indicators to generate structured output data, construct a mapping relationship between input data and output data, and train a forward prediction model and a generative model based on the mapping relationship. The forward prediction model is used to predict performance indicators through design parameters, and the generative model is used to predict design parameters through performance indicators.

[0155] The optimization module 203 is used to receive different target performance indicators as inputs, generate a set of candidate design parameters using the trained generative model, use the set of candidate design parameters as the initial population, and predict the corresponding set of predicted performance indicators through the forward prediction model, determine the Pareto dominance relationship of the set of predicted performance indicators, and perform multi-objective optimization on the initial population based on the Pareto dominance relationship to obtain the set of target design parameters.

[0156] Using the aforementioned device, a bidirectional prediction model is first constructed using historical antenna design data. This data preprocessing and model training mechanism effectively overcomes the limitations of traditional trial-and-error methods, eliminating the need for extensive repetitive simulations in the optimization process. Then, by combining the high-quality initial solutions output by the generated model with a multi-objective evolutionary algorithm, the system can efficiently search for the globally optimal solution set within the design space, while automatically balancing conflicts among multiple dimensions such as radiation characteristics and bandwidth. Compared to traditional methods that are prone to getting trapped in local optima and have high computational costs, this invention significantly accelerates the performance evaluation process using a surrogate model and provides diverse non-dominated solution choices through Pareto front analysis, allowing for flexible trade-offs between different design objectives based on actual needs. This method is particularly suitable for MIMO antenna designs that require consideration of complex electromagnetic performance, and its modular architecture also demonstrates the potential for expansion into other electronic device optimization scenarios.

[0157] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The steps of the provided MIMO antenna design optimization method.

[0158] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various operations. The processor reads the corresponding computer program from the non-volatile memory into the memory and then executes it to achieve the above-mentioned functions. Figure 1 The steps of the provided MIMO antenna design optimization method.

[0159] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0161] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0162] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0163] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the patent of the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for designing and optimizing a MIMO antenna, comprising: The method includes: Obtain the historical design parameters and corresponding performance indicators of the MIMO antenna; Historical design parameters are encoded to generate structured input data, performance indicators are normalized to generate structured output data, a mapping relationship between input and output data is constructed, and a forward prediction model and a generative model are trained based on the mapping relationship; the forward prediction model is used to predict performance indicators through design parameters; the generative model is used to predict design parameters through performance indicators. Different target performance indicators are received as inputs, and a candidate design parameter set is generated using a trained generative model. The candidate design parameter set is used as the initial population, and the corresponding predicted performance indicator set is predicted by a forward prediction model. The Pareto dominance relationship of the predicted performance indicator set is determined, and multi-objective optimization is performed on the initial population based on the Pareto dominance relationship to obtain the target design parameter set. The forward prediction model and the generative model trained based on the mapping relationship include: By training a CNN or GNN, a nonlinear mapping from design parameters to performance metrics is learned, resulting in a forward prediction model; by training a VAE or GAN, an inverse mapping from performance metrics to design parameters is learned, resulting in a generative model; wherein, the CNN is used to process the geometric structure of image / voxel encoding and extract spatial features; the GNN is used to process the array parameters of graph structure encoding and explicitly model the coupling effect between units. The process of generating a candidate design parameter set using the trained generative model includes: Input the target performance index into the conditional generative adversarial network (GAN) to directly generate candidate design parameters; Alternatively, sampling optimization can be performed using the latent space of a variational autoencoder (VAE): The target performance is mapped to the target region of the potential space, potential vectors are sampled in the target region, and candidate design parameters are generated based on the potential vectors through the VAE decoder.

2. The method of claim 1, wherein, The historical design parameters include geometric parameters and material parameters. The process of encoding the historical design parameters to generate structured input data and normalizing the performance indicators to generate structured output data includes: Vector encoding, image / voxel matrix encoding, or graph structure encoding are used for geometric parameters; Material parameters can be input directly using numerical values ​​or encoded using unique thermal methods. Frequency-related performance indicators are sampled as fixed-length vectors.

3. The method of claim 2, wherein, The graph structure encoding includes: The antenna elements in the MIMO array are treated as graph nodes. The node features include the element's own parameters, and the edge features include the estimated values ​​of the element spacing or coupling strength.

4. A device for design optimization of a MIMO antenna, characterized in that The device includes: The acquisition module is used to acquire the historical design parameters and corresponding performance indicators of the MIMO antenna. The design parameters include geometric parameters, material parameters, and feed network parameters; the performance indicators include S-parameters, radiation characteristics, MIMO characteristics, and bandwidth. A construction module is used to encode historical design parameters to generate structured input data, normalize performance indicators to generate structured output data, construct a mapping relationship between input and output data, and train a forward prediction model and a generative model based on the mapping relationship. The forward prediction model is used to predict performance indicators through design parameters; the generative model is used to predict design parameters through performance indicators. The training of the forward prediction model and generative model based on the mapping relationship includes: training a CNN or GNN to learn a nonlinear mapping from design parameters to performance indicators to obtain the forward prediction model; and training a VAE or GAN to learn an inverse mapping from performance indicators to design parameters to obtain the generative model. The CNN is used to process the geometric structure of image / voxel encoding and extract spatial features; the GNN is used to process the array parameters of graph structure encoding and explicitly model the coupling effect between units. An optimization module is used to receive different target performance indicators as inputs, generate a set of candidate design parameters using a trained generative model, use the set of candidate design parameters as an initial population, and predict the corresponding set of predicted performance indicators using a forward prediction model. The Pareto dominance of the predicted performance indicator set is determined, and multi-objective optimization is performed on the initial population based on the Pareto dominance to obtain the target design parameter set. Specifically, generating the set of candidate design parameters using the trained generative model includes: inputting the target performance indicators into a Generative Adversarial Network (GAN) to directly generate candidate design parameters; or, sampling optimization using the latent space of a Variational Autoencoder (VAE): mapping the target performance to a target region in the latent space, sampling latent vectors in the target region, and generating candidate design parameters based on the latent vectors using a VAE decoder.

5. A computer readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 3.

6. A computer device, comprising: The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 3.

Images