CN-122021519-A - Electromagnetic simulation method for numerical green function antenna radiation field based on neural network

CN122021519ACN 122021519 ACN122021519 ACN 122021519ACN-122021519-A

Abstract

The invention discloses a numerical green function antenna radiation field electromagnetic simulation method based on a neural network, and belongs to the technical field of electromagnetic simulation. The method comprises the steps of taking a free space green function analytic solution as a physical priori, constructing a weighted approximation network by using a multi-layer perceptron MLP, obtaining a numerical green function preliminary approximation solution by weighted approximation, calculating a residual error of the preliminary approximation solution and the numerical green function, performing high-precision fitting on the residual error through a KAN network, superposing MLP network output and KAN network output to obtain a high-precision numerical green function prediction result, and solving antenna radiation/scattering field distribution by combining an electromagnetic scattering field calculation formula. The method can realize efficient and high-precision solving of the antenna radiation field and the target electromagnetic scattering characteristic under a complex scene.

Inventors

HU JUN
XUE XIAOHAN
JIANG MING
ZHAO RAN
YANG XIONG
ZONG XIANZHENG

Assignees

电子科技大学

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (7)

1. The electromagnetic simulation method of the numerical green function antenna radiation field based on the neural network is characterized by comprising the following steps: step 1, setting a solving scene of an antenna radiation field, wherein the solving scene contains a fixed target, calculating by a numerical method to obtain a part of source point positions and real numerical green function values corresponding to the field points in the setting solving scene, and constructing a first training sample data set by using the source point position information, the field point position information and the corresponding real numerical green function values; Step2, constructing a weighted approximation network, namely training the weighted approximation network by using a free space green function analytic solution as a physical priori constraint and using a first training sample data set to obtain a trained weighted approximation network; Step 3, calculating residual errors between the initial approximate solution of the numerical green function and the real numerical green function value, and constructing a second training sample data set by using the fixed target position information, the source point position information, the field point position information and the corresponding residual errors; Step 4, constructing a residual fitting network, training the residual fitting network by using a second training sample data set to obtain a trained residual fitting network; the trained weighted approximation network and residual fitting network are cascaded to be used as a double-layer neural network structure, wherein the numerical green function prediction result output by the double-layer neural network structure is the superposition of the numerical green function preliminary approximation solution output by the weighted approximation network and the residual output by the residual fitting network; the method comprises the steps of adopting a mean square error to verify the precision of a double-layer neural network structure, and judging that the double-layer neural network structure meets the precision requirement if the mean square error between a numerical green function prediction result and a real numerical green function value is smaller than a set threshold value; And 5, inputting the position information of the radiation source of the test antenna, the position information of the test field point, the fixed target position information and the free space green function analysis solution into a trained double-layer neural network structure to obtain a numerical green function prediction result corresponding to the test antenna and the test field point, and solving the field distribution of the test target in the radiation field of the test antenna based on the numerical green function prediction result.
2. The electromagnetic simulation method of a numerical green function antenna radiation field based on a neural network according to claim 1, wherein in step 1, the training samples in the first training sample data set include source point position information and field point position information, and the sample label is a real numerical green function value corresponding to the feature parameters.
3. The electromagnetic simulation method of the numerical green function antenna radiation field based on the neural network according to claim 1, wherein in the step 2, the weighted approximation network selects a multi-layer perceptron, inputs the information of the relative position between the source point and the field point and the free space green function analytic solution, and outputs the information as the numerical green function value.
4. The electromagnetic simulation method of a numerical green function antenna radiation field based on a neural network according to claim 1, wherein in step 2, a free space green function analytic solution in a two-dimensional scene is provided Expressed as: Wherein, the And J is an imaginary symbol, and k is a wave number; Is a second class zero-order hank function; Representing the distance between the field point and the source point in the two-dimensional plane.
5. The electromagnetic simulation method of the numerical green function antenna radiation field based on the neural network according to claim 1, wherein in the step 2, the free space green function analytic solution in the three-dimensional scene is Expressed as: Wherein, the And The method comprises the steps of respectively obtaining position vectors of field points and source points in a three-dimensional space, wherein j is an imaginary symbol, k is a wave number, e represents a natural base, and R represents the distance between the field points and the source points in the three-dimensional space.
6. The neural network-based numerical green function antenna radiation field electromagnetic simulation method according to claim 1, wherein in step 3, the residual training samples in the second training sample data set include fixed target position information, source point position information and field point position information, and the sample label is a residual corresponding to the characteristic parameters.
7. The electromagnetic simulation method of the numerical green function antenna radiation field based on the neural network according to claim 1, wherein in the step 4, a KAN network is selected as a residual fitting network, and the KAN network is input as fixed target position information, source point position information and field point position information, and the KAN network is output as residual; After training, the numerical green function preliminary approximation solution output by the weighted approximation network is calculated Fitting the residual to the residual Superposing to obtain the final numerical green function prediction result : Mean square error Expressed as: Wherein, the And Respectively representing the normalized numerical green function prediction result and the actual numerical green function value corresponding to the ith training sample, wherein i is more than or equal to 1 and less than or equal to n, and n is the total number of the training samples.

Description

Electromagnetic simulation method for numerical green function antenna radiation field based on neural network Technical Field The invention belongs to the technical field of electromagnetic simulation, and particularly relates to a numerical green function antenna radiation field electromagnetic simulation method based on a neural network. Background The antenna is used as a core device for electromagnetic energy radiation and reception, and the radiation field characteristics of the antenna directly determine the overall performance of a wireless communication system, a radar detection system and the like. In practical engineering application, the antenna works in complex scenes containing metal targets, dielectric barriers and multi-antenna coupling, the antenna radiation field and the target electromagnetic scattering characteristics in the scenes are accurately solved, and the method has important research significance and engineering value for optimizing the design of the antenna structure, evaluating the system performance, formulating the electromagnetic interference suppression strategy and the like. However, electromagnetic problems in complex scenes relate to multi-scale structures, strong coupling effects and nonlinear electromagnetic responses, and the traditional solving method faces the bottleneck that accuracy and efficiency are difficult to be compatible. The method is one of core methods for solving the antenna radiation field and electromagnetic scattering problems, the core principle is that the electromagnetic problems of a complex scene are converted into solution of special green functions of the scene, electromagnetic coupling relations between source points (antenna excitation sources) and field points are accurately represented through the green functions, and the radiation field distribution of the antenna in the complex scene can be directly solved by means of convolution operation of the numerical green functions and the excitation sources on the premise of knowing the current/charge distribution of the antenna excitation sources. According to the method, through the electromagnetic interaction of the characterization radiation source and the scene target, repeated calculation in the multi-source and multi-field point solving scene is effectively reduced, the obvious calculation efficiency advantage is achieved, the problems of high calculation complexity and long time consumption still exist in the traditional numerical Green function solving process, and particularly in a high-frequency and complex multi-scale engineering scene, the requirements of actual application on instantaneity and high efficiency are difficult to meet, and the large-scale popularization and landing of the method in the engineering field are greatly limited. Along with the rapid development of artificial intelligence technology, the application of the artificial intelligence technology in the field of electromagnetic simulation provides a new thought for solving the bottleneck, and gradually becomes a research hot spot. At present, artificial intelligence methods such as neural networks and the like are widely applied to tasks such as antenna optimization design, electromagnetic scattering rapid prediction, electromagnetic parameter inversion and the like, and the core thought is to learn an internal physical rule of an electromagnetic problem through a model so as to replace or accelerate a traditional numerical calculation process. The solution of the numerical green function needs to traverse all source-field relations in the calculation domain, is a work consuming huge calculation resources, and can realize the efficient acceleration of the solution process by means of artificial intelligence technology. The Chinese patent application with the application number 202510273609.4 discloses a numerical green function acceleration solving method of a three-dimensional moving target group, the Chinese patent application with the application number 202510768307.4 discloses a neural network acceleration simulation method of a broadband target group radar scattering field, the numerical green function is fitted through a single structure network in the prior art, only simple physical prior information is introduced into a part of schemes, the core physical rules of electromagnetic problems are not fused deeply enough, the generalization capability and solving precision of a model are limited, and the high-precision solving requirement of an antenna radiation field under a complex scene is difficult to adapt. In summary, a hybrid modeling method of depth fusion physical prior and intelligent acceleration technology is needed in the current engineering field, so that efficient and high-precision solution of an antenna radiation field in a complex scene is realized, and the technical problems of low efficiency, and insufficient precision and generalization of the traditional numerical green function method are s