Antenna surface current prediction method based on deep learning algorithm

Abstract

The application belongs to the technical field of antennas, and particularly relates to an antenna surface current prediction method based on a deep learning algorithm. A forward neural network based on a U-Net framework is adopted, a pixelized antenna structure is taken as input, and an antenna preset resolution surface current distribution result under a preset working frequency is output. A training data set is obtained by calculation of an electromagnetic simulation software, and network training is completed by taking a mean square error as a loss function. The method can quickly predict the antenna surface current distribution, provides a physical basis for obtaining an antenna electromagnetic field, a response curve and a directional diagram, greatly reduces simulation and design costs, and is suitable for efficient analysis and design of patch antennas.

Description

Technical Field

[0001] This invention belongs to the field of antenna technology, specifically relating to a method for predicting antenna surface current based on deep learning algorithms. Background Technology

[0002] In modern radio frequency and wireless communication systems, antennas are key components for transmitting and receiving electromagnetic waves, and their electromagnetic performance directly affects the overall system performance. Parameters such as the antenna's electric field distribution, magnetic field distribution, reflection coefficient S11 response curve, and radiation pattern are all closely related to the current distribution on the antenna conductor surface. Therefore, surface current distribution is a crucial physical quantity connecting the antenna structure and its electromagnetic radiation performance, and is of great significance for antenna design and performance analysis.

[0003] Traditional antenna design methods typically rely on design experience, involving pre-setting topologies and continuously adjusting geometric parameters, followed by iterative calculations and optimizations using full-wave electromagnetic simulation software to obtain an antenna structure that meets design specifications. This approach is computationally expensive, has a long design cycle, and is inefficient under complex structures or large parameter spaces. With the development of artificial intelligence, research has attempted to use machine learning or deep learning methods to establish mapping relationships between antenna structures and performance indicators such as S-parameters and radiation patterns, or to combine optimization algorithms to perform structural parameter search, thereby improving design efficiency to some extent.

[0004] However, existing methods mostly focus on the direct mapping between structural and electromagnetic performance results, lacking modeling of key intermediate physical quantities in the antenna electromagnetic radiation mechanism. Since surface current distribution can reflect the actual operating state of the antenna and can be further used to deduce electric field, magnetic field, and radiation characteristics, achieving rapid and accurate prediction of it would help improve the physical interpretability of the model and design efficiency. Therefore, it is necessary to propose a deep learning-based antenna surface current prediction method to replace the time-consuming calculation process in traditional electromagnetic simulation and achieve rapid acquisition of antenna surface current distribution. Summary of the Invention

[0005] The purpose of this invention is to provide an antenna surface current prediction method based on deep learning algorithms, so as to achieve fast and accurate prediction of the surface current distribution of a given antenna structure, thereby replacing the problem of complex calculation and long time consumption in traditional full-wave electromagnetic simulation, and improving the efficiency of antenna analysis and design.

[0006] The antenna surface current prediction method proposed in this invention uses a forward neural network based on the U-Net framework as the deep learning model to construct a mapping relationship between the antenna structure and its current distribution, thereby achieving rapid prediction of the current distribution of an unknown antenna structure. The method includes the following steps (taking a pixelated patch antenna as an example):

[0007] (1) First, construct a pixelated patch antenna model and establish the corresponding antenna structure data;

[0008] (2) Then, the surface current distribution corresponding to the pixelated patch antenna is calculated using electromagnetic simulation software at a preset operating frequency to construct a supervised learning dataset; the dataset is then divided into a training set, a validation set and a test set to train the forward neural network; during the training process, the optimization objective is to minimize the error between the predicted current distribution and the actual current distribution.

[0009] (3) Finally, the antenna structure whose surface current distribution is to be predicted is input into the trained forward neural network, and the network outputs the corresponding surface current distribution result.

[0010] Furthermore:

[0011] The pixelated patch antenna has the following basic structure: the upper layer is a 16×16 pixelated metal patch, and the patch structure is guaranteed to have four-connected continuity through morphological operations; the middle layer is a dielectric layer; the back of the lower dielectric substrate is completely covered with metal; the feed point position is randomly selected in the 16×16 pixels, and the corresponding pixel is forced to be metal.

[0012] The pixelated patch antenna structure serves as the input to the forward neural network, specifically a two-dimensional array representing the patch structure, where element 0 represents a blank area, 1 represents a metal area, and 2 represents the feeding position; the output result is the surface current distribution result with a resolution of 32×32, that is, the current value at a uniformly distributed position on the antenna surface, and each current component includes a real part and an imaginary part; the surface current distribution is the patch surface current distribution obtained at the preset operating frequency.

[0013] The forward neural network adopts a U-Net network structure, specifically including: an input layer, an encoder downsampling module, a feature bottleneck layer, a decoder upsampling module, a skip connection module, and an output layer; wherein:

[0014] The input layer is used to receive a two-dimensional array input characterizing the antenna structure; in the two-dimensional array, element 0 represents a blank area, 1 represents a metal area, and 2 represents the feed position.

[0015] The encoder downsampling module consists of multi-level convolutional units and downsampling units, which are used to extract the local geometric features, connectivity features and feed position features of the input antenna structure step by step, and to perform dimensionality reduction and compression on the feature map to expand the network receptive field and obtain deep semantic features.

[0016] The feature bottleneck layer is set between the encoder downsampling module and the decoder upsampling module to further characterize the downsampled high-dimensional features in order to learn the nonlinear mapping relationship between the antenna structure and the surface current distribution.

[0017] The decoder upsampling module consists of multi-level upsampling units and convolutional units, which are used to recover and reconstruct the deep features of the bottleneck layer output step by step to obtain a surface current distribution feature map corresponding to the target output resolution.

[0018] The jump connection module is used to splice or fuse the shallow features of the corresponding scale in the encoder with the deep features of the corresponding scale in the decoder, so as to retain the boundary information, local topology information and feed-side details of the pixelated patch antenna and improve the surface current distribution prediction accuracy.

[0019] The output layer is used to output the antenna surface current distribution result at a set resolution; the surface current distribution result is the current component value at a uniform position on the antenna surface, and each component includes a real part and an imaginary part.

[0020] Furthermore, the dataset required for training the forward neural network is obtained by electromagnetic simulation software. Specifically, the following steps are taken: First, pixelated patch antenna structure samples that satisfy four-connectivity continuity are randomly generated; then, the electromagnetic simulation software is used to perform full-wave electromagnetic calculations on each antenna structure at a preset operating frequency to obtain surface current distribution data at uniform sampling positions on the corresponding patch surface; the surface current distribution data serves as a supervised learning label, including the values ​​of each component of the current at each sampling position on the antenna surface, and each component contains real and imaginary part information.

[0021] The antenna structure data obtained from the electromagnetic simulation and the corresponding surface current distribution data together constitute a supervised learning dataset. The input data is a two-dimensional array representing the antenna structure, and the output data is the corresponding surface current distribution result. The size of the dataset is adjusted according to the spatial complexity of the antenna structure design, the accuracy requirements of the prediction target, and the convergence of the network training, to ensure that the neural network can fully learn the mapping relationship between the antenna structure and the surface current distribution. The obtained dataset is divided into a training set, a validation set, and a test set. The training set is used for network parameter learning and weight updates; the validation set is used to monitor the trend of the loss function during training, determine the network convergence, and evaluate whether overfitting has occurred; the test set is used to evaluate the generalization prediction ability of the trained neural network for unknown antenna structure samples.

[0022] Furthermore, the forward neural network training aims to minimize the loss function value between the predicted surface current distribution and the actual current distribution, thereby establishing a stable and accurate nonlinear mapping relationship between the antenna structure input and the surface current distribution output. The network's fitting ability to known samples is assessed by examining the trend of the loss function on the training set; the generalization ability and overfitting phenomenon during training are assessed by examining the trend of the loss function on the validation set; and the network's prediction performance on the new test dataset (test set) is evaluated to determine the current prediction ability of the trained network for unknown antenna structure samples. The forward neural network training is considered complete when the loss functions on both the training and validation sets tend to converge stably, and good prediction accuracy is maintained on the test set. If the network performance after training is poor, the following hyperparameters are adjusted to improve model performance: training dataset size, number of training epochs, loss function form, learning rate, batch size, number of convolutional layers, number of feature channels, and network structure parameters.

[0023] Furthermore, during the training of the forward neural network, the Mean Squared Error (MSE) function is used as the loss function to measure the error between the surface current distribution predicted by the network at the preset operating frequency and the actual surface current distribution obtained from electromagnetic simulation; the loss function is defined at the preset operating frequency as follows:

[0024]

[0025] Where, 𝑁 represents the number of surface current sampling points. No. The actual current value at each sampling point The current value is predicted by the neural network; by minimizing the loss function, the error between the predicted result and the actual result is minimized.

[0026] Furthermore, the training cycle of the forward neural network can be set to 500 epochs; in each epoch, the prediction result is calculated through forward propagation, and the network weight parameters are updated through backpropagation algorithm based on the loss function.

[0027] Furthermore, the predicted antenna surface current distribution can be used as an intermediate physical quantity for antenna electromagnetic performance analysis. It can be further used to solve for the corresponding antenna's electric field distribution, magnetic field distribution, reflection coefficient S11 response curve, radiation pattern, as well as electromagnetic performance parameters such as gain, main lobe direction, and radiation efficiency, thereby achieving a rapid evaluation of the overall antenna performance.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] By replacing traditional full-wave electromagnetic simulation with a deep learning model, rapid prediction of antenna surface current distribution is achieved, significantly reducing computation time costs. Introducing surface current distribution as an intermediate physical quantity provides better physical interpretability compared to methods that directly establish a mapping relationship between structure and performance. This method can effectively improve the efficiency of complex antenna structure analysis and design, and is suitable for intelligent antenna design scenarios. Furthermore, this method can be extended to different types of antenna structures, demonstrating good versatility and application prospects. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the basic structure of a pixelated patch antenna.

[0031] Figure 2 This is the overall architecture of the antenna current prediction method of the present invention.

[0032] Figure 3 This is a schematic diagram of a forward neural network model based on the U-Net structure.

[0033] Figure 4 (a) shows the surface current distribution obtained by the software solution. Figure (4)b is a schematic diagram of the prediction results of the antenna surface current distribution by the method of the present invention.

[0034] In the diagram, the numbers are: 1 is the power supply location, 2 is the pixelated metal patch, 3 is the dielectric layer, and 4 is the metal ground plane. Detailed Implementation

[0035] The antenna surface current prediction method based on deep learning algorithms proposed in this invention will be further described below with reference to the accompanying drawings. This embodiment uses a pixelated patch antenna as an example for illustration, but the method is not limited to this structural form.

[0036] Figure 1 This is a schematic diagram of the pixelated patch antenna structure selected in this embodiment. The antenna includes an upper metal patch, an intermediate dielectric layer, and a lower metal ground plane. The upper metal patch is discretely represented using a pixelated method, divided into a 16×16 grid; the intermediate layer is the dielectric substrate; and the lower layer is a complete metal ground plane. The feed point is selected within the patch grid and encoded as a specific marker in the structural representation. To ensure the physical realizability of the antenna structure, the pixelated patch structure undergoes morphological processing to satisfy the four-connectivity continuity constraint.

[0037] In this embodiment, the antenna structure is represented as a 16×16 two-dimensional array as the input to the neural network, where matrix element 0 represents an empty area, 1 represents a metallic area, and 2 represents the feed position. This two-dimensional array can completely describe the antenna's geometric topology and feed information.

[0038] During the dataset construction process, pixelated patch antenna samples that satisfy connectivity constraints are first randomly generated within a preset structural space; then, electromagnetic simulation software is used to perform full-wave electromagnetic calculations on each antenna structure at a preset operating frequency, where the frequency is set to 10 GHz, to obtain the corresponding surface current distribution data.

[0039] Subsequently, uniform spatial sampling is performed on the patch surface to generate discrete current data with a resolution of 32×32; the current value at each sampling location includes the real and imaginary parts of the corresponding directional component. A supervised learning dataset is constructed by mapping the antenna structure data one-to-one with the corresponding current distribution data.

[0040] In this embodiment, a total of 1000 sets of sample data were generated and divided into a training set, a validation set, and a test set, with 900 sets in the training set, 50 sets in the validation set, and 50 sets in the test set. The training set is used for learning network parameters, the validation set is used for convergence monitoring and model tuning during the training process, and the test set is used to evaluate the model's generalization ability to unknown samples.

[0041] like Figure 2 The diagram shown is an overall flowchart of the method of this invention. The new antenna structure is input into the trained model, and the network outputs the corresponding surface current distribution results.

[0042] like Figure 3 The diagram shows a forward neural network model based on the U-Net structure used in this embodiment. The network includes an input layer, an encoder downsampling module, a feature bottleneck layer, a decoder upsampling module, and an output layer. The encoder extracts multi-scale features of the input structure through multi-level convolution and downsampling operations; the decoder gradually restores spatial resolution through upsampling operations; and the encoder and decoder fuse features through skip connections to preserve structural boundary information and improve prediction accuracy. In this embodiment, the convolutional layers preferably use 3×3 convolutional kernels, the downsampling method preferably uses max pooling (with a stride of 2), and the upsampling method uses deconvolution.

[0043] In this embodiment, the neural network input is a 16×16 antenna structure matrix, and the output is a 32×32 resolution surface current distribution result at the preset operating frequency. The output result is the current value at each sampling position on the patch surface, including real and imaginary part information.

[0044] During network training, the Mean Squared Error (MSE) function is used as the loss function, and its expression is:

[0045]

[0046] in, This represents the number of surface current sampling points. No. The actual current value at each sampling point The value is the current value predicted by the neural network. The loss function is defined at a preset operating frequency and is used to measure the error between the prediction result and the simulation result.

[0047] The network parameters are continuously updated through iterative training, causing the loss function to gradually converge. When the loss function on the training and validation sets stabilizes and the prediction results on the test set meet the accuracy requirements, the network training is considered complete.

[0048] After training, the new pixelated patch antenna structure is input into the forward neural network, which can quickly output the corresponding surface current distribution results. Compared with traditional full-wave electromagnetic simulation methods, this invention eliminates the need to repeatedly solve electromagnetic equations, significantly improving computational efficiency.

[0049] like Figure 4 The diagram shown illustrates the predicted current distribution on the antenna surface in this embodiment. A comparison of the predicted results with the electromagnetic simulation results demonstrates that the method of this invention can accurately reproduce the current distribution characteristics on the antenna surface.

[0050] Based on the predicted surface current distribution, the electric field distribution and magnetic field distribution of the corresponding antenna can be further solved, and electromagnetic performance parameters such as the reflection coefficient S11 response curve and radiation pattern can be derived, thereby realizing a rapid evaluation of the overall performance of the antenna.

[0051] The results of this embodiment demonstrate the effectiveness of the method of the present invention in predicting antenna surface current.

Claims

1. A method for predicting antenna surface current based on a deep learning algorithm, characterized in that, A forward neural network based on the U-Net framework is used as a deep learning model to construct a mapping relationship between the antenna structure and its current distribution, enabling rapid prediction of the current distribution of an unknown antenna structure. For pixelated patch antennas, the following steps are included: (1) First, construct a pixelated patch antenna model and establish the corresponding antenna structure data; (2) Then, the surface current distribution corresponding to the pixelated patch antenna is calculated using electromagnetic simulation software at a preset operating frequency to construct a supervised learning dataset; the dataset is then divided into a training set, a validation set and a test set to train the forward neural network. During training, the optimization objective is to minimize the error between the predicted current distribution and the actual current distribution. (3) Finally, the antenna structure whose surface current distribution is to be predicted is input into the trained forward neural network, and the network outputs the corresponding surface current distribution result. The pixelated patch antenna has the following basic structure: the upper layer is a 16×16 pixelated metal patch, and the patch structure is guaranteed to have four-connected continuity through morphological operations; the middle layer is a dielectric layer; the back of the lower dielectric substrate is completely covered with metal; the feed point position is randomly selected in the 16×16 pixels, and the corresponding pixel is forced to be metal. The pixelated patch antenna structure serves as the input to the forward neural network, specifically a two-dimensional array representing the patch structure, where element 0 represents a blank area, 1 represents a metal area, and 2 represents the feeding position; the output result is the surface current distribution result with a resolution of 32×32, that is, the current value at a uniformly distributed position on the antenna surface, and each current component includes a real part and an imaginary part; the surface current distribution is the patch surface current distribution obtained at the preset operating frequency.

2. The antenna surface current prediction method according to claim 1, characterized in that, The forward neural network adopts a U-Net network structure, specifically including: an input layer, an encoder downsampling module, a feature bottleneck layer, a decoder upsampling module, a skip connection module, and an output layer; wherein: The input layer is used to receive a two-dimensional array input characterizing the antenna structure; in the two-dimensional array, element 0 represents a blank area, 1 represents a metal area, and 2 represents the feed position. The encoder downsampling module consists of multi-level convolutional units and downsampling units, which are used to extract the local geometric features, connectivity features and feed position features of the input antenna structure step by step, and to perform dimensionality reduction and compression on the feature map to expand the network receptive field and obtain deep semantic features. The feature bottleneck layer is set between the encoder downsampling module and the decoder upsampling module to further characterize the downsampled high-dimensional features in order to learn the nonlinear mapping relationship between the antenna structure and the surface current distribution. The decoder upsampling module consists of multi-level upsampling units and convolutional units, which are used to recover and reconstruct the deep features of the bottleneck layer output step by step to obtain a surface current distribution feature map corresponding to the target output resolution. The jump connection module is used to splice or fuse the shallow features of the corresponding scale in the encoder with the deep features of the corresponding scale in the decoder, so as to retain the boundary information, local topology information and feed-side details of the pixelated patch antenna and improve the surface current distribution prediction accuracy. The output layer is used to output the antenna surface current distribution result at a set resolution; the surface current distribution result is the current component value at a uniform position on the antenna surface, and each component includes a real part and an imaginary part.

3. The antenna surface current prediction method according to claim 2, characterized in that, The dataset required for training the forward neural network is obtained by electromagnetic simulation software. Specifically, the following steps are taken: First, pixelated patch antenna structure samples that satisfy four-connectivity continuity are randomly generated; then, the electromagnetic simulation software is used to perform full-wave electromagnetic calculations on each antenna structure at a preset operating frequency to obtain surface current distribution data at uniform sampling positions on the corresponding patch surface; the surface current distribution data serves as a supervised learning label, including the values ​​of each component of the current at each sampling position on the antenna surface, and each component contains real and imaginary part information. The antenna structure data obtained from electromagnetic simulation and the corresponding surface current distribution data together constitute a supervised learning dataset. The input data is a two-dimensional array representing the antenna structure, and the output data is the corresponding surface current distribution result. The size of the dataset is adjusted according to the spatial complexity of the antenna structure design, the accuracy requirements of the prediction target, and the convergence of the network training, to ensure that the neural network can fully learn the mapping relationship between the antenna structure and the surface current distribution. The obtained dataset is divided into a training set, a validation set, and a test set. The training set is used for network parameter learning and weight updates; the validation set is used to monitor the trend of the loss function during training, determine the network convergence, and evaluate whether overfitting has occurred; the test set is used to evaluate the generalization prediction ability of the trained neural network for unknown antenna structure samples.

4. The antenna surface current prediction method according to claim 3, characterized in that, The goal of training a forward neural network is to minimize the loss function value between the predicted surface current distribution and the actual current distribution, thereby establishing a stable and accurate nonlinear mapping relationship between the antenna structure input and the surface current distribution output. The network's fitting ability to known samples is assessed by examining the trend of the loss function on the training set; the generalization ability and overfitting during training are assessed by examining the trend of the loss function on the validation set; and the network's prediction performance on the new test set (test dataset) is evaluated to determine its current prediction ability for unknown antenna structure samples after training. The forward neural network training is considered complete when the loss functions on both the training and validation sets tend to converge stably, and good prediction accuracy is maintained on the test set. If the network performance after training is poor, the following hyperparameters are adjusted to improve model performance: training dataset size, number of training epochs, loss function form, learning rate, batch size, number of convolutional layers, number of feature channels, and network structure parameters.

5. The antenna surface current prediction method according to claim 4, characterized in that, During the training of the forward neural network, the mean square error function (MSE) is used as the loss function to measure the error between the surface current distribution predicted by the network at the preset operating frequency and the actual surface current distribution obtained from electromagnetic simulation. The loss function is defined at the preset operating frequency as follows: Where, 𝑁 represents the number of surface current sampling points. No. The actual current value at each sampling point The current value is predicted by the neural network; by minimizing the loss function, the error between the predicted result and the actual result is minimized.

6. The antenna surface current prediction method according to claim 5, characterized in that, The training cycle of the forward neural network is set to 500 epochs; in each epoch, the prediction result is calculated through forward propagation, and the network weight parameters are updated through backpropagation algorithm based on the loss function.

7. The antenna surface current prediction method according to claim 6, characterized in that, The predicted antenna surface current distribution serves as an intermediate physical quantity for antenna electromagnetic performance analysis. It can be further used to solve for the corresponding antenna's electric field distribution, magnetic field distribution, reflection coefficient S11 response curve, radiation pattern, as well as electromagnetic performance parameters such as gain, main lobe direction, and radiation efficiency, thereby enabling a rapid evaluation of the overall antenna performance.

Images