Antenna performance prediction and optimization method based on physical information guided deep learning

Abstract

This invention specifically relates to a method for antenna performance prediction and optimization based on physical information-guided deep learning, belonging to the fields of antenna design and artificial intelligence technology. The method includes: acquiring antenna geometric parameters and corresponding full-wave simulation S11 curves, constructing a sample dataset; extracting physical features such as resonance depth, resonant frequency, and upper and lower bandwidth limits from the S11 curves; constructing a hierarchical neural network model, including a geometric parameter encoding sub-network, a physical feature prediction sub-network, and a frequency domain response prediction sub-network; employing a multi-task joint training method, using the S11 curve prediction error and the physical feature prediction error together to form a loss function for optimization; using the trained model as a surrogate model, combining it with an intelligent optimization algorithm to quickly search for antenna geometric parameters, and outputting the optimal design that meets the performance requirements of the target frequency band. This invention improves the prediction accuracy and physical consistency of the antenna frequency domain response, significantly reduces simulation costs, and enhances antenna design efficiency.

Description

Technical Field

[0001] This invention relates to the fields of antenna design and artificial intelligence technology, and to a method for predicting and optimizing the performance of miniaturized broadband antennas based on physical information-guided deep learning. Specifically, it relates to an antenna frequency domain response prediction method that integrates physical feature constraints and an intelligent optimization method for antenna geometric parameters based on the prediction model. Background Technology

[0002] With the evolution of wireless communication technology towards 5G and 6G and the explosive development of the Internet of Things (IoT), increasingly stringent requirements are being placed on antenna performance. Electrically Small Antennas (ESA) are antennas with a maximum size less than λ / 2π. Due to their size being much smaller than the operating wavelength, they have wide applications in modern mobile communications, IoT devices, wearable electronics, aerospace, and biomedicine. However, the design of ESAs is a complex multi-parameter, multi-objective optimization problem. According to the Chu-Harrington limit theory, ESAs are subject to fundamental physical limitations, exhibiting inherent defects such as high quality factor, narrow bandwidth, low radiation efficiency, and low gain. Furthermore, their performance (such as return loss S11, gain, and bandwidth) is extremely sensitive to changes in structural parameters, and the parameter space often contains numerous local optima, posing a significant challenge to traditional design methods.

[0003] Traditional antenna design primarily relies on the following methods: first, analytical design methods based on empirical formulas, which are computationally simple but have limited accuracy and struggle to handle complex structures; second, numerical optimization methods based on electromagnetic simulation, which offer high accuracy but involve enormous computational demands, with a single simulation taking minutes to hours; and third, exhaustive search methods based on parameter scanning, which offer high reliability but are extremely inefficient, resulting in design cycles lasting months. These traditional methods can no longer meet the demands of rapid antenna design in modern times.

[0004] In recent years, intelligent optimization algorithms such as Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and Differential Evolution (DE) have been widely used in antenna design. These methods primarily rely on electromagnetic simulation software (such as CST and HFSS) combined with optimization algorithms (such as Genetic Algorithm and Particle Swarm Optimization). However, these methods suffer from the following key problems: First, slow convergence speed, easily getting trapped in local optima; second, parameter settings heavily rely on experience, lacking adaptive adjustment capabilities; third, difficulty in maintaining population diversity, easily leading to premature convergence; fourth, a single selection strategy limits the algorithm's global search capability; and fifth, a lack of effective constraint handling mechanisms.

[0005] In addition, this type of method also faces the following inherent drawbacks in practical applications: (1) High computational cost: Every parameter modification requires a complete full-wave electromagnetic simulation. For electrically small antennas, a single simulation may take several minutes to several hours even on a high-performance computer. Effective global optimization often requires tens of thousands of iterations, with a total time consumption of several days or even weeks, resulting in extremely low design efficiency.

[0006] (2) Curse of Dimension: Antenna performance is a common function of multiple geometric parameters (such as length, width, height, feed position, etc.). As the number of optimization variables increases, the parameter space grows exponentially, and traditional methods struggle to find the global optimum in a finite amount of time.

[0007] (3) Reliance on expert experience: The selection of initial parameters and the formulation of optimization strategies rely heavily on the designer's experience and intuition. There is a lack of universal automated design process, which is not conducive to the rapid promotion and iteration of technology.

[0008] In recent years, machine learning technology has been introduced into the field of antenna design, enabling performance prediction and assisted optimization by constructing a nonlinear mapping relationship between antenna geometric parameters and electromagnetic performance indicators. However, most existing machine learning-based antenna modeling methods typically treat the antenna frequency domain response (such as the S-parameter curve) as a single regression objective, ignoring the physical characteristics such as resonant frequency, resonant depth, and bandwidth inherent in the frequency domain response. This results in insufficient physical consistency of the model and limited generalization ability. Furthermore, in the parameter optimization stage, without the guidance of physical constraints, the optimization algorithm is prone to getting trapped in local optima or producing infeasible designs.

[0009] Therefore, there is an urgent need for a technical solution that can explicitly introduce physical feature information, balance prediction accuracy and physical consistency, and can be used for efficient optimization of antenna parameters. Summary of the Invention

[0010] This invention provides a method for antenna performance prediction and optimization based on physical information-guided deep learning. It aims to solve the technical problems in the design of electrically small antennas, such as the contradiction between accuracy and efficiency in traditional methods, the tendency of intelligent optimization algorithms to reach local optima and high computational cost, the lack of physical constraints in existing machine learning modeling leading to poor generalization, and the over-reliance on expert experience in overall design, making it difficult to achieve rapid and automated optimization.

[0011] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0012] According to a first aspect of the present invention, a method for antenna performance prediction and optimization based on physically-informed guided deep learning is provided, the method comprising: Step 1, Antenna Sample Data Acquisition and Preprocessing: Define the antenna geometry using a parameterized approach, acquire multiple sets of antenna geometric parameter samples, and perform frequency domain simulation on each set of geometric parameters using a full-wave electromagnetic simulation tool to obtain the corresponding frequency domain reflection parameter S11 curve; normalize the geometric parameter samples and the frequency domain response S11 curve samples respectively to form a sample dataset; Step 2, Physical Feature Extraction: For each S11 frequency domain response curve, extract key physical feature parameters and construct a physical feature vector; Step 3, Construct a frequency domain prediction model guided by physical information: Construct a hierarchical neural network model, which includes at least: A geometric parameter encoding subnetwork is used to perform feature mapping on antenna geometric parameters; The physical feature prediction subnetwork is used to predict the corresponding physical feature parameters based on the geometric parameters. A frequency domain response prediction subnetwork is used to predict the complete S11 frequency domain curve under the joint constraints of geometric parameter features and the prediction results of the physical features. Step 4, Multi-task joint training: The neural network model is trained using a multi-task learning approach. The prediction error of the S11 frequency domain curve and the prediction error of the physical feature parameters are used to form a joint loss function. Through joint optimization, the model can meet the physical consistency constraint while ensuring the accuracy of frequency domain prediction. Step 5, Antenna parameter optimization based on prediction model: After completing model training, the neural network model is used as a proxy model for antenna performance. Combined with intelligent optimization algorithm, the antenna geometric parameters are searched and optimized to meet the optimization target of the reflection coefficient being lower than a preset threshold in the target frequency band. The optimized combination of antenna geometric parameters is then output.

[0013] In some exemplary embodiments, the physical feature vector extracted in step two includes the resonance depth, the main resonant frequency, and the upper and lower bandwidth limits frequency under a preset threshold. The minimum reflection coefficient value in the frequency domain response, i.e., the resonance depth; The frequency point corresponding to the minimum reflection coefficient, i.e., the main resonant frequency; , : respectively satisfy The lower and upper limits of the frequency range, where This is the preset reflection coefficient threshold.

[0014] In some exemplary embodiments, the input to the physical feature prediction subnetwork is a geometric parameter vector, and the physical feature vector is directly predicted through three fully connected layers.

[0015] In some exemplary embodiments, the frequency domain response prediction subnetwork concatenates geometric parameters with predicted physical features and outputs a complete S11 frequency domain response curve through four fully connected layers. The first three fully connected layers use the ReLU activation function, and a Dropout layer is introduced to prevent overfitting.

[0016] In some exemplary embodiments, the joint loss function in step four is expressed as:

[0017] in, The mean square error of the S11 frequency domain curve prediction. The mean square error of the predicted physical characteristic parameters. Hyperparameters used to balance the weights of the two.

[0018] In some exemplary embodiments, the intelligent optimization algorithm used in step five is an improved genetic algorithm, which includes the following steps: Initialize the population, setting the population size and number of generations; The fitness function is defined as minimizing the maximum S11 value within the target frequency band. Set constraints that require the S11 value to be lower than a preset threshold throughout the entire target frequency band; Evolutionary operations are performed using a simulated binary crossover operator and a polynomial mutation operator; Iterate and evolve until the convergence condition is met, and output the optimal combination of geometric parameters.

[0019] In some exemplary embodiments, after optimization is completed, the optimized antenna geometric parameter combination is verified by full-wave electromagnetic simulation, and the consistency between the predicted results and the simulation results is compared to verify the reliability of the optimization results.

[0020] According to a second aspect of the present invention, an antenna performance prediction and optimization system based on physically-informed deep learning is provided, the system comprising: The data acquisition and preprocessing module is used to acquire antenna geometric parameter samples and their corresponding S11 frequency domain response curves, and perform normalization processing. The physical feature extraction module is used to extract physical features from the S11 curve, including resonance depth, resonant frequency, and upper and lower limits of bandwidth. The model building and training module is used to build hierarchical neural network models and train them using a multi-task joint training method. The parameter optimization module uses a trained neural network model as a surrogate model and combines it with an intelligent optimization algorithm to optimize and search for the antenna's geometric parameters, and outputs the optimal combination of geometric parameters.

[0021] According to a third aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the antenna performance prediction and optimization method based on physical information-guided deep learning as described in the first aspect.

[0022] According to a fourth aspect of the present invention, a computer program product is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, it implements the antenna performance prediction and optimization method based on physical information-guided deep learning as described in the first aspect.

[0023] According to a fifth aspect of the present invention, an electronic device is provided, comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to implement the antenna performance prediction and optimization method based on physical information-guided deep learning as described in the first aspect by executing the executable instructions.

[0024] The antenna performance prediction and optimization method based on physical information-guided deep learning provided by the embodiments of the present invention has at least the following beneficial effects compared with the prior art: by introducing frequency domain physical features as intermediate constraints, the physical consistency and generalization ability of the frequency domain response prediction model are improved; by adopting a multi-task joint training mechanism, the accurate prediction of key physical indicators is achieved while ensuring the overall prediction accuracy of the S11 curve; by using the trained prediction model as a surrogate model for parameter optimization, the dependence on full-wave electromagnetic simulation is greatly reduced, and the computational cost is significantly reduced; the method has strong versatility and can be extended to the design and optimization of different types of antennas and other radio frequency devices.

[0025] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0026] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0027] Figure 1 This is a schematic diagram of the overall process of the method of the present invention; Figure 2 A schematic diagram of a frequency domain response prediction model guided by physical information; Figure 3 This is a schematic diagram of the antenna parameter optimization process based on a prediction model. Figure 4 This is a schematic diagram of the antenna geometry; Figure 5 A comparison chart of the predicted S11 curve from the prediction model and the actual electromagnetic simulation S11 curve; Figure 6 A comparison chart of the S11 curves for artificial optimization simulation, model optimization simulation, and model optimization prediction. Detailed Implementation

[0028] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0029] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0030] This invention proposes solutions to many pain points in the design of miniaturized broadband electrically small antennas and in existing antenna performance prediction and parameter optimization techniques. The core technical problems addressed include: 1. Optimization challenges arising from inherent design flaws of electrically small antennas: Electrically small antennas are limited by the Chu-Harrington limit theory, resulting in inherent problems such as high quality factor and narrow bandwidth. Furthermore, their performance is extremely sensitive to changes in geometric parameters, and there are a large number of local optima in the parameter space. Traditional design methods are difficult to achieve efficient optimization of multiple parameters and multiple objectives.

[0031] 2. The contradiction between efficiency and accuracy in traditional antenna design methods: analytical methods based on empirical formulas have limited accuracy and cannot handle complex antenna structures; numerical methods based on electromagnetic simulation and exhaustive methods based on parameter scanning involve huge amounts of computation and are extremely time-consuming, with a single simulation taking several minutes to several hours, and global optimization taking even several days to several months, which is difficult to meet the needs of rapid antenna design in modern times.

[0032] 3. Application defects of traditional intelligent optimization algorithms: When intelligent algorithms such as genetic algorithms and particle swarm optimization are combined with electromagnetic simulation, there are problems such as slow convergence speed, easy to get trapped in local optima, parameter setting depends on experience, difficulty in maintaining population diversity, and lack of effective constraint processing mechanism. At the same time, they also face inherent defects such as high computational cost, curse of dimensionality, and over-reliance on expert experience.

[0033] 4. Limitations of existing machine learning antenna modeling methods: Existing machine learning-based antenna modeling treats frequency domain response as a single regression target, ignoring physical characteristics such as resonant frequency, resonant depth, and bandwidth, resulting in insufficient physical consistency of the model and limited generalization ability; moreover, the lack of physical constraints during the optimization stage makes it easy to produce local optima or infeasible designs.

[0034] 5. Insufficient automation and universality in antenna design: The selection of initial parameters and the formulation of optimization strategies in existing methods rely heavily on the designer's experience and intuition. There is no universally applicable automated design process, which is not conducive to the rapid promotion and iteration of antenna design technology.

[0035] To address the shortcomings and deficiencies of existing technologies, this example implementation provides an antenna performance prediction and optimization method based on physical information-guided deep learning. By introducing frequency domain physical feature constraints into the deep learning model, high-precision prediction of the antenna S-parameter curve is achieved. Furthermore, based on this prediction model, intelligent optimization of antenna geometric parameters is realized, thereby significantly reducing simulation costs and improving design efficiency.

[0036] refer to Figure 1 As shown, the method of the present invention may specifically include the following steps: Step 1, Antenna Sample Data Acquisition and Preprocessing: Define the antenna geometry using a parameterized approach. Let the antenna geometric parameter vector be: (1.1) in, The antenna's first Several geometric parameters were obtained. Multiple sets of antenna geometric parameter samples were acquired, and a full-wave electromagnetic simulation tool was used to simulate the antenna within a preset frequency range. For each set of geometric parameters, frequency domain simulation is performed to obtain the corresponding frequency domain reflection parameter S11 curve: (1.2) in, These are discrete frequency sampling points. The geometric parameter samples and the frequency domain response S11 curve samples are normalized to form a sample dataset for model training.

[0037] Step 2, Physical Feature Extraction: For each S11 frequency domain response curve, key feature parameters are extracted from a physical perspective to construct a physical feature vector. (1.3) in: This is the minimum reflection coefficient value (resonance depth) in the frequency domain response. The frequency point corresponding to the minimum reflection coefficient (main resonant frequency); , They respectively satisfy The lower and upper limits of the frequency range, Set a preset threshold (e.g., -10 dB).

[0038] In this way, the original high-dimensional frequency domain response is mapped into a low-dimensional feature representation with clear physical semantics.

[0039] Step 3: Construct a frequency domain prediction model guided by physical information: Construct a hierarchical neural network model, which includes at least: First, a geometric parameter encoding subnetwork, used for feature mapping of antenna geometric parameters; Second, a physical feature prediction subnetwork, used for predicting corresponding physical feature parameters based on geometric parameters; These two parts are used to establish a nonlinear mapping relationship between geometric parameters and physical features. (1.4) in, This represents a physical feature prediction network. Its network parameters.

[0040] Third, the frequency domain response prediction subnetwork is used to predict the complete S11 frequency domain curve under the joint constraints of the geometric parameter features and the prediction results of the physical features: (1.5) The input to this subnetwork is a vector of geometric parameters. and predict physical eigenvectors splicing, Represented as a frequency domain prediction network function, It is its set of network parameters.

[0041] Step 4, Multi-task Joint Training: The neural network model is trained using a multi-task learning approach, and the prediction error of the S11 frequency domain curve and the prediction error of the physical feature parameters are used together to form a joint loss function. (1.6) (1.7) (1.8) in, The weighting coefficients for the physical feature loss are used. By jointly optimizing the above loss functions, the model can satisfy the physical consistency constraint while ensuring the accuracy of frequency domain prediction.

[0042] Step 5, Antenna parameter optimization based on the prediction model: After completing the model training, the neural network model is used as a surrogate model for antenna performance. Combined with the intelligent optimization algorithm, the antenna geometric parameters are searched and optimized. The optimization objective is to ensure that the reflection coefficient in the target frequency band is lower than a preset threshold. The optimized combination of antenna geometric parameters is then output.

[0043] Example 1 This embodiment takes a vertically small electrically small monopole antenna as an example. The target operating frequency band is 100MHz to 400MHz, and the return loss S11 is required to be less than -10dB in this frequency band.

[0044] Step 1, Data Preparation and Preprocessing: Using full-wave simulation software, a dataset containing 2832 different antenna design samples was generated through parametric scanning, including 2265 training samples and 567 test samples. For example... Figure 4 A schematic diagram of the antenna geometry is shown. Each sample of the antenna contains six key geometric parameters (d1, d3, Rresi, topload, trh, tri1). For each sample, the S11 parameters at 301 equally spaced frequency points within the 100-400MHz range are simulated and calculated. The input geometric parameter matrix X and the output S11 matrix Y are normalized using MinMaxScaler, scaling them to the [0, 1] interval. Furthermore, four key physical features are extracted from the S11 curve: minimum S11 value (resonance depth), resonant frequency, lower limit frequency of the -10dB bandwidth, and upper limit frequency of the -10dB bandwidth, forming a physical feature vector p, which is also normalized.

[0045] Step 2, Construction and training of a deep agent model guided by physical information: using attached... Figure 2 The physical information-guided hierarchical network structure is shown. This network consists of two parts: (1) a physical feature extraction subnetwork: the input is 6-dimensional geometric parameters, which are directly predicted through three fully connected layers (128→64→4 nodes); (2) a frequency domain response prediction subnetwork: the geometric parameters are concatenated with the predicted physical features (a total of 10-dimensional features), and the complete S11 frequency domain response curve is output through four fully connected layers (512→512→256→301 nodes). Each hidden layer uses the ReLU activation function and adds a Dropout layer with a Dropout rate of 0.1 to prevent overfitting. A multi-task learning strategy is adopted, and the joint loss function is:

[0046] in The mean square error of the S11 prediction. λ represents the mean square error of the physical feature prediction, and is set to 0.2.

[0047] Training was performed using the Adam optimizer with a batch size of 32 for a total of 150 epochs. The training process was completed on a CPU and took 155.2 seconds. After training, the model achieved excellent performance on the test set: a prediction accuracy of 91.32% within a ±1 dB tolerance range. Physical feature prediction performance was as follows: the mean absolute error for minimum S11 prediction was 2.90 dB, the mean absolute error for resonant frequency prediction was 8.57 MHz, and the mean absolute errors for lower and upper bandwidth predictions were 14.02 MHz and 18.87 MHz, respectively. Figure 5 The comparison between the predicted S11 curve from the prediction model and the actual electromagnetic simulation S11 curve clearly shows that the model has high accuracy and can be used for the prediction of S-parameters.

[0048] Step 3, Antenna Parameter Optimization Based on Genetic Algorithm: The trained physical information-guided network is used as a high-performance surrogate model and embedded in the optimization process. The optimization problem is defined as: within the range of values ​​for six geometric parameters (determined by the distribution of training data and prior knowledge), find a set of parameters that minimizes the maximum S11 value in the 100-400MHz target frequency band, and constrains the S11 value of the entire frequency band to be below -10dB. An improved genetic algorithm (GA) based on the Pymoo library is used. The algorithm parameters are set as follows: population size is set to 100, and the maximum number of generations is 100. Simulated binary crossover (SBX, probability 0.9, distribution exponent 15) and multinomial mutation (PM, distribution exponent 20) operators are used.

[0049] The fitness function (F) is: .

[0050] Constraints for: ,Require .

[0051] The maximum S11 value (i.e. fitness) of a population decreases continuously with the number of generations, and tends to converge after about the 30th generation.

[0052] Step 4, Result Verification and Analysis: After 100 generations of evolution (10,000 evaluations in total), the genetic algorithm converged, finding a set of optimal geometric parameters as shown in the table below. The surrogate model predicts that the maximum S11 of this design in the 100-400MHz frequency band is -11.2764 dB, the minimum S11 is -23.1340 dB, and the average S11 is -14.9534 dB, with all 301 frequency points having an S11 below -10dB (100% satisfaction). The predicted key physical characteristics are: minimum S11 of -25.1030dB, resonant frequency of 213.9142MHz, and a -10dB bandwidth covering 88.07 MHz to 393.26 MHz (bandwidth approximately 305.19 MHz). These results are attached. Figure 6 As shown in the right figure, the S11 curve is smooth and meets the required depth throughout the entire target frequency band. After verification through full-wave electromagnetic simulation, the model predictions and actual simulation values ​​show a high degree of consistency, significantly exceeding the simulation values ​​obtained through manual parameter tuning and optimization. The total optimization time was extremely short: the surrogate model training (a one-time investment) took approximately several minutes, while the optimization search based on the surrogate model took only about 1.39 seconds, demonstrating high efficiency.

[0053]

[0054] If the traditional method of combining full-wave simulation with optimization algorithms is used, assuming a single simulation takes 5 minutes, completing 10,000 evaluations of the same scale would require approximately 50,000 minutes (about 34.7 days). The method of this invention reduces the optimization search time to the second level, improving overall efficiency by tens of thousands of times, while ensuring high accuracy and reliability of the design. This fully demonstrates the effectiveness and significant engineering application value of the physical information-guided modeling method and intelligent optimization process proposed in this invention.

[0055] This invention uses physical information to guide deep learning to build a prediction model and combines it with intelligent optimization algorithms to achieve antenna performance prediction and parameter optimization. Compared with existing technologies, it achieves several significant benefits, including: 1. Improve the physical consistency and generalization ability of the prediction model: By explicitly introducing frequency domain physical features as intermediate constraints, the high-dimensional frequency domain response is mapped to low-dimensional features with clear physical semantics. This solves the problem of traditional machine learning models ignoring physical information, making the model prediction results more consistent with the actual electromagnetic physical characteristics of the antenna, and improving the generalization ability of the model under different antenna parameter scenarios.

[0056] 2. Ensure high accuracy in antenna performance prediction: A multi-task joint training mechanism is adopted, which incorporates the prediction error of S11 frequency domain curve and the prediction error of physical characteristic parameters into the loss function. While ensuring the overall prediction accuracy of S11 curve, accurate prediction of key physical indicators such as resonance depth, resonance frequency, and bandwidth is achieved. The measured prediction accuracy can reach 91.32% within the ±1dB tolerance range.

[0057] 3. Significantly reduced computational costs and improved design efficiency: The trained deep learning prediction model is used as a proxy model for antenna performance, replacing traditional full-wave electromagnetic simulation for parameter optimization, completely eliminating the reliance on a large number of repetitive simulations. Traditional methods require approximately 34.7 days to complete optimization and evaluation of the same scale, while the optimization search of this invention takes only seconds, improving overall design efficiency by tens of thousands of times. Moreover, model training is a one-time investment, taking only a few minutes.

[0058] 4. Achieve efficient global optimization of antenna parameters: Combined with an improved intelligent optimization algorithm for parameter search, based on a surrogate model constrained by physical information, it effectively avoids the problems of traditional optimization algorithms being prone to getting trapped in local optima and premature convergence. It can quickly find the global optimal solution in a multi-parameter space, and the optimization results meet the performance requirements of the target frequency band.

[0059] 5. Reduce reliance on expert experience and achieve automated design: This invention constructs a standardized and streamlined antenna performance prediction and optimization system. From sample data processing and model building and training to parameter optimization, the entire process is automated. It does not require designers to rely on their experience and intuition to select initial parameters or formulate optimization strategies, thereby improving the universality of the technology and facilitating rapid promotion and iteration.

[0060] 6. Strong versatility and extended application value: The method of this invention is not designed for a specific type of antenna, but can be extended to different types of electrically small antennas. It can also be transferred to the performance prediction and geometric parameter optimization of other radio frequency devices, providing a general technical solution for the rapid design of radio frequency devices.

[0061] 7. The optimization results are both reliable and practical: Through full-wave electromagnetic simulation verification, the predicted values ​​after model optimization are highly consistent with the actual simulation values, and the optimized antenna parameters enable the antenna performance indicators in the target frequency band to fully meet the design requirements. Compared with the results of manual parameter tuning optimization, it is better and has practical engineering application value.

[0062] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0063] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0064] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is defined only by the appended claims.

Claims

1. A method for antenna performance prediction and optimization based on physically-informed deep learning, characterized in that, The method includes: Step 1, Antenna Sample Data Acquisition and Preprocessing: Define the antenna geometry using a parameterized approach, acquire multiple sets of antenna geometric parameter samples, and perform frequency domain simulation on each set of geometric parameters using a full-wave electromagnetic simulation tool to obtain the corresponding frequency domain reflection parameter S11 curve; normalize the geometric parameter samples and the frequency domain response S11 curve samples respectively to form a sample dataset; Step 2, Physical Feature Extraction: For each S11 frequency domain response curve, extract key physical feature parameters and construct a physical feature vector; Step 3, Construct a frequency domain prediction model guided by physical information: Construct a hierarchical neural network model, which includes at least: A geometric parameter encoding subnetwork is used to perform feature mapping on antenna geometric parameters; The physical feature prediction subnetwork is used to predict the corresponding physical feature parameters based on the geometric parameters. A frequency domain response prediction subnetwork is used to predict the complete S11 frequency domain curve under the joint constraints of geometric parameter features and the prediction results of the physical features. Step 4, Multi-task joint training: The neural network model is trained using a multi-task learning approach. The prediction error of the S11 frequency domain curve and the prediction error of the physical feature parameters are used to form a joint loss function. Through joint optimization, the model can meet the physical consistency constraint while ensuring the accuracy of frequency domain prediction. Step 5, Antenna parameter optimization based on prediction model: After completing model training, the neural network model is used as a proxy model for antenna performance. Combined with intelligent optimization algorithm, the antenna geometric parameters are searched and optimized to meet the optimization target of the reflection coefficient being lower than a preset threshold in the target frequency band. The optimized combination of antenna geometric parameters is then output.

2. The method according to claim 1, characterized in that, The physical feature vector extracted in step two includes resonance depth, main resonant frequency, and upper and lower bandwidth limits under a preset threshold. The minimum reflection coefficient value in the frequency domain response, i.e., the resonance depth; The frequency point corresponding to the minimum reflection coefficient, i.e., the main resonant frequency; , : respectively satisfy The lower and upper limits of the frequency range, where This is the preset reflection coefficient threshold.

3. The method according to claim 1, characterized in that, The physical feature prediction subnetwork takes a geometric parameter vector as input and directly predicts the physical feature vector through three fully connected layers.

4. The method according to claim 1, characterized in that, The frequency domain response prediction subnetwork concatenates geometric parameters with predicted physical features and outputs a complete S11 frequency domain response curve through four fully connected layers. The first three fully connected layers use the ReLU activation function and introduce a Dropout layer to prevent overfitting.

5. The method according to claim 1, characterized in that, The joint loss function in step four is expressed as follows: in, The mean square error of the S11 frequency domain curve prediction. The mean square error of the predicted physical characteristic parameters. Hyperparameters used to balance the weights of the two.

6. The method according to claim 1, characterized in that, The intelligent optimization algorithm used in step five is an improved genetic algorithm, which includes the following steps: Initialize the population, setting the population size and number of generations; The fitness function is defined as minimizing the maximum S11 value within the target frequency band. Set constraints that require the S11 value to be lower than a preset threshold throughout the entire target frequency band; Evolutionary operations are performed using a simulated binary crossover operator and a polynomial mutation operator; Iterate and evolve until the convergence condition is met, and output the optimal combination of geometric parameters.

7. The method according to claim 1, characterized in that, After optimization, the optimized antenna geometric parameter combination is verified by full-wave electromagnetic simulation. The consistency between the predicted results and the simulation results is compared to verify the reliability of the optimization results.

8. An antenna performance prediction and optimization system based on physically-informed deep learning, characterized in that, include: The data acquisition and preprocessing module is used to acquire antenna geometric parameter samples and their corresponding S11 frequency domain response curves, and perform normalization processing. The physical feature extraction module is used to extract physical features from the S11 curve, including resonance depth, resonant frequency, and upper and lower limits of bandwidth. The model building and training module is used to build hierarchical neural network models and train them using a multi-task joint training method. The parameter optimization module uses a trained neural network model as a surrogate model and combines it with an intelligent optimization algorithm to optimize and search for the antenna's geometric parameters, and outputs the optimal combination of geometric parameters.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the antenna performance prediction and optimization method based on physical information-guided deep learning as described in any one of claims 1 to 7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the antenna performance prediction and optimization method based on physical information-guided deep learning as described in any one of claims 1 to 7.

Images