A method for predicting the scattering field of a variable-curvature frequency selective surface unit
By combining neural network models with the infinitesimal dipole method, the problems of scarce computational resources and repetitive modeling of scattering fields of surface units with variable curvature frequency selection are solved, and efficient and accurate scattering field prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2025-02-27
- Publication Date
- 2026-06-09
AI Technical Summary
Traditional algorithms suffer from scarce computational resources and high costs when calculating the scattering field of surface units with variable curvature frequencies, and the efficiency of repeated modeling and simulation analysis during the design process is low.
A neural network model combined with the infinitesimal dipole method is used to generate a sample set through logarithmic sampling, establish a frequency-selective surface unit model library, build a long short-term memory deep neural network, and train the model to predict the scattering field.
It improves computational efficiency and accuracy, solves the problems of high computational resource requirements and repetitive modeling in traditional methods, and expands the application of surface scattering calculation with variable curvature frequency selection.
Smart Images
Figure CN120068649B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of radome technology, specifically relating to a method for predicting the scattering field of a frequency-selective surface element with variable curvature, which can be used to assist in the calculation and analysis of scattering fields of large-scale frequency-selective surfaces. Background Technology
[0002] Frequency-selective surfaces (FMS) are large-scale structures composed of identical frequency-selective cells arranged in a two-dimensional periodic pattern. Due to their spatial filtering properties, they are widely used in military and wireless communication systems. However, due to their complex periodic structure and electrically large size, traditional low-frequency algorithms (such as the method of moments) often face challenges of limited computational resources and high computational costs when processing these surfaces. To address the problem of calculating the scattering characteristics of FMS, some researchers have proposed using the frequency-selective cells as basic units for array calculations, with significant improvements. In this method, the calculation of the scattered field of the frequency-selective cells becomes crucial.
[0003] For structures like frequency-selective elements, which possess mixed metallic and dielectric properties, two main methods exist for calculating their scattering fields: numerical methods and equivalent methods. Numerical methods, such as the method of moments, the finite-difference time-domain method, and the finite element method, often suffer from excessive computational costs due to the large number of meshes. The infinitesimal dipole model equivalent method (IDM equivalent method), however, differs from traditional mesh-based methods. The IDM equivalent method uses a set of infinitesimal dipoles to represent the current distribution on the model surface. This method can ignore the object's external shape, offering significant advantages for scattering calculations on frequency-selective elements. International research has demonstrated the superiority of this algorithm. Mikki, for example, introduced Quantum Particle Swarm Optimization (QPSO) into the equivalent process. Through numerous trials of the equivalent model, the QPSO optimization algorithm reduced the degrees of freedom of each dipole from ten in the genetic algorithm to seven. Yang optimized the seven parameters into two by restricting the position and orientation of a set of IDMs, and then optimized the parameters using near-field data, ultimately successfully predicting the scattering field results.
[0004] In practical applications, the frequency selective unit is initially designed as a planar unit. However, when the frequency selective unit is applied to a large-scale frequency selective surface (such as a radome), the shape of the frequency selective unit changes from planar to irregular. This causes many problems for designers, and the repetitive modeling and simulation analysis greatly reduces work efficiency.
[0005] Therefore, a prediction method for the scattering field of frequency-selective surface units with variable curvature is proposed, which is of great significance in engineering applications of large-scale frequency-selective surface scattering calculations. Summary of the Invention
[0006] To address the aforementioned problems, this invention proposes a method for predicting the scattering field of frequency-selective surface units with variable curvature. This method, by building a neural network model, can accurately calculate the scattering field of frequency-selective surface units with arbitrary curvature characteristics, which is of great significance for large-scale frequency-selective surface scattering calculations.
[0007] The present invention is achieved through the following technical solution.
[0008] This invention provides a method for predicting the scattering field of a frequency-selective surface element with variable curvature, comprising the following steps:
[0009] Logarithmically sample the curvature of the frequency-selective surface unit and convert it to the original spatial curvature to generate a curvature sampling point sample set.
[0010] A frequency-selective surface unit model library is established based on the curvature sampling point sample set;
[0011] The established frequency-selective surface element model library was simulated and solved to obtain near-field data of frequency-selective surface elements with different curvatures.
[0012] Select the surface unit model library based on the frequency and solve for the infinitesimal dipole position vector matrix;
[0013] Based on the near-field data of the surface unit and the infinitesimal dipole position vector matrix selected according to different curvature frequencies, the dipole moment in the infinitesimal dipole group is solved to obtain the dipole moment dataset of different units.
[0014] Based on the dipole moment dataset, a long short-term memory deep neural network model was constructed.
[0015] Based on the long short-term memory deep neural network model, adjust the hyperparameters of the neural network model and train the neural network model;
[0016] Based on the trained neural network model, the dipole moment is predicted and combined with the infinitesimal dipole group position vector matrix to solve the scattering field of the frequency-selective surface unit, thus completing the prediction of the scattering field of the frequency-selective surface unit.
[0017] As a preferred method, logarithmic sampling is performed on the curvature of the frequency-selective surface unit, and sampling is performed on the arc at the logarithmic scale. The curvature at the logarithmic scale is restored to the original spatial curvature sampling points to generate a sample set.
[0018] As a preferred option, a frequency-selective surface element model library is established, including:
[0019] Let the two orthogonal planes XOZ and YOZ be the primary and secondary circular arc surfaces, respectively. The geometry within the primary circular arc surface is a closed surface formed by two circular arcs and two line segments. The secondary circular arc surface is composed of a single circular arc. The sweeping section within the primary circular arc surface is scanned along the circular arc trajectory within the secondary circular arc surface to generate the geometry.
[0020] As a preferred approach, simulations are performed on the established frequency-selective surface element model library, including:
[0021] The established frequency-selective surface element solid model was used in a full-wave simulation software to set observation points, plane wave form, and solution options to solve for near-field data of frequency-selective surface elements with different curvatures.
[0022] As a preferred method, a surface element model library is selected based on the frequency, and the infinitesimal dipole position vector matrix is solved, including:
[0023] Discretize the circular arcs in the YOZ plane into nodes to obtain the position information of the dipoles;
[0024] The dipole position coordinates on the inner arc of the YOZ plane are obtained by rotating the dipole coordinates along the X-axis;
[0025] Three orthogonally arranged dipoles are placed at each position, and the information and position information are arranged into a position vector matrix.
[0026] As a preferred method, the dipole moments in the infinitesimal dipole group are solved to obtain a dataset of dipole moments for different elements, including:
[0027] Combining the free space and the Losglin function, based on the infinitesimal dipole model, the expression for the scattering near field is represented. The common factor in the scattering field expression of the infinitesimal dipole model is extracted and rewritten as the product of the dipole moment and the position vector matrix of the infinitesimal dipole.
[0028] Perform singular value decomposition on the position vector matrix;
[0029] Solve for the dipole moments of the infinitesimal dipole group so that the field calculation results of the infinitesimal dipole group are in agreement with the full-wave simulation results, and organize the dipole moments of different units in the sample set into a dipole moment dataset.
[0030] As a preferred option, a long short-term memory deep neural network model is constructed, including:
[0031] The Long Short-Term Memory (LSTM) deep neural network model consists of an input layer and hidden layers. The input layer has two time steps, with each time step having a feature dimension of 1, and 50 samples are used for training each time. The hidden layer has two layers, each with 256 neurons. The dropout ratio between LSTM layers is 0.3 to prevent overfitting. The output layer has 150 neurons, which is the same dimension as the final regression target.
[0032] Preferably, adjusting the hyperparameters of the neural network model and training the neural network model includes:
[0033] The data in the dataset is first normalized. Since the data in the dataset consists of complex numbers, it is divided into real part datasets and imaginary part datasets for separate training. Then, it is divided into training set, validation set and test set in a ratio of 7:2:1. The curvature features of each sample are used as labels to annotate the dataset.
[0034] The neural network model is trained with a learning rate of 0.0001 during training, which decays to 0.85 every 65 steps. If the loss value does not decrease within 50 steps using early stopping logic, training is stopped.
[0035] Preferably, the scattering field of the frequency-selective surface unit is solved, including:
[0036] The trained neural network model is used as a database. The predicted data is denormalized and then combined with the infinitesimal dipole position vector matrix to form a predicted dipole group.
[0037] Substituting the predicted dipole group into the formula for calculating the scattering pattern of the frequency-selective unit, the scattering field of the frequency-selective surface unit is obtained.
[0038] The present invention, by adopting the above technical solution, has the following beneficial effects:
[0039] 1. This invention employs the infinitesimal dipole model method, which can calculate the scattering field results of the frequency-selective unit, improving the equivalent accuracy while solving the problems of long equivalent time and high computational resource requirements in the traditional infinitesimal dipole model equivalent method.
[0040] 2. This invention proposes a method for reconstructing a frequency-selective surface element model. The required model is generated by parameterization through curve sweeping, which overcomes the difficulty in obtaining array elements in a variable curvature frequency-selective surface and expands the application of the infinitesimal dipole model method in scattering calculations of variable curvature frequency-selective surfaces.
[0041] 3. This invention establishes a long short-term memory deep neural network and an end-to-end network model between the curvature and dipole moment of frequency-selective surface units, avoiding the repetitive modeling and equivalence work in the infinitesimal dipole model method, and improving the design efficiency of frequency-selective surfaces. Attached Figure Description
[0042] Figure 1 This is a flowchart of a method for predicting the scattering field of a variable curvature frequency-selective surface unit according to the present invention;
[0043] Figure 2 This is the modeling process of the frequency selection unit in this invention;
[0044] Figure 3 This is the equivalent idea of the infinitesimal dipole model equivalent method in this invention;
[0045] Figure 4 Basic information about the frequency selection unit examples used in this invention;
[0046] Figure 5 The results are the prediction comparison of the real part of the dipole moment of the dipole group in frequency-selective unit 1;
[0047] Figure 6 The results are the prediction comparison of the real part of the dipole moment of the dipole group in frequency-selective unit 2;
[0048] Figure 7 The results of the prediction comparison of the imaginary part of the dipole moment of the dipole group in frequency-selective unit 1;
[0049] Figure 8 The results of the prediction comparison of the imaginary part of the dipole moment of the dipole group in frequency-selective unit 2;
[0050] Figure 9 The scattering pattern of the frequency-selective unit 1 is compared with the simulation results of the commercial software FEKO using the present invention;
[0051] Figure 10 The scattering pattern of the frequency-selective unit 2 is compared with the simulation results of the commercial software FEKO. Detailed Implementation
[0052] The invention will now be described in further detail with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the invention in any way.
[0053] Reference Figure 1 This invention provides a flowchart of a method for predicting the scattering field of a variable curvature frequency-selective surface unit. The specific steps are as follows:
[0054] Step 1: Generate a sample set based on the logarithmic sampling principle.
[0055] Based on the logarithmic sampling principle, the curvature of the frequency-selective surface unit is logarithmically sampled and converted back to the original spatial curvature to generate a curvature sampling point sample set, including the following steps:
[0056] (1a) For frequency-selective surface element curvature sampled on a logarithmic scale, the curvature z on the logarithmic scale is:
[0057]
[0058] Where, x min and x max, respectively, represent the lower and upper limits of the curvature sampling interval, z represents the curvature sampling points on the logarithmic scale, N represents the number of sampling points, and i represents the sampling sequence number.
[0059] (1b) The curvature at the logarithmic scale is restored to the original spatial curvature sampling points, where the curvature x in the original space is:
[0060] x = exp(z)
[0061] Where x is the curvature sampling point mapped back to the original space on a logarithmic scale.
[0062] Step 2: Establish a variable curvature frequency-selective unit model library
[0063] Establishing a frequency-selective element model of arbitrary shape based on different curvatures includes the following steps:
[0064] Let the two orthogonal planes XOZ and YOZ be the principal and secondary circular arc surfaces, respectively. The geometry within the principal circular arc surface consists of a closed surface formed by two circular arcs and two line segments, with the radii of curvature R of the arcs being R. x and R x -h; The secondary circular arc surface consists of a circular arc with a radius of curvature R. y The geometry is generated by sweeping the cross section within the main circular arc surface along the circular arc trajectory within the secondary circular arc surface.
[0065] Step 3: Perform near-field simulation analysis on the unit.
[0066] The established frequency-selective surface element solid model library was used to set the observation points, plane wave form, and solution options in the full-wave simulation software. Simulations were then performed to solve the near-field data of frequency-selective surface elements with different curvatures.
[0067] Step 4: Generate the dipole position vector matrix
[0068] Based on the frequency-selective unit model library established in step 2, the dipole moments in the infinitesimal dipole group are solved, including the following steps:
[0069] (4a) Discretize the circular arc in the YOZ plane into nodes to obtain the position information of the dipole; the coordinates of the discrete points of the circular arc are as follows:
[0070]
[0071] in, Let r be the coordinates of a discrete point on the arc; r be the curvature of the arc; θ be the coordinates of a point on the arc. i For discrete angles of the circular arc.
[0072] (4b) Rotate the dipole coordinates on the inner arc of the YOZ plane along the X-axis to obtain the dipole position coordinates on the frequency-selective element surface.
[0073]
[0074] in, The dipole coordinates of the frequency-selective cell surface; φ is the rotation matrix along the X-axis; j The discrete angle of the inner arc of the secondary arc.
[0075] (4c) Place three dipoles at each position, and place the three dipoles orthogonally to organize the pointing information and position information into a position vector matrix C.
[0076] C = [α,β,x,y,z]
[0077] Where α is the angle between the dipole and the X-axis, and the dipoles pointing to the X, Y, and Z axes are set to 0°, 90°, and 90° respectively; β is the angle between the dipole and the Y-axis, and the dipoles pointing to the X, Y, and Z axes are set to 90°, 0°, and 90° respectively; x, y, and z are the position coordinates of the dipole in free space.
[0078] Step 5: Solve for the dipole moments in the system of infinitesimal dipoles.
[0079] Based on the near-field data results of the surface elements obtained in step 3 (different curvature frequencies), and the dipole position vector matrix obtained in step 4, the dipole moments in the infinitesimal dipole group are solved, including the following steps:
[0080] (5a) The expression for the free space non-Green's function is derived as follows:
[0081]
[0082] in, The identity matrix is represented by k; k is the propagation coefficient. Let be the scalar Green's function; r and r' are the position vectors of the near-field observation point and the infinitesimal dipole, respectively.
[0083] (5b) Combining the free space and the Losgreen function, the expression for the scattered field is based on the infinitesimal dipole model (IDM):
[0084]
[0085] Where j is the complex unit; ω is the angular frequency; μ is the free space permeability; α i β i γ i Let x, y, and z be the angles between the i-th infinitesimal dipole and the X, Y, and Z axes, respectively; x, y, and z are the unit vectors along the X, Y, and Z axes, respectively; M i N is the dipole moment of the infinitesimal dipole; i is the dipole number, N d Let E be the total number of dipoles, and E be the scattering near-field result obtained using dipoles. It is a dyadic Green's function.
[0086] (5c) Extract the common factor of the scattering field expression of the infinitesimal dipole model and rewrite it as the product of the dipole moment and the position vector matrix.
[0087]
[0088] Among them, E mask The scattering near-field results are from the full-wave simulation software; C i It is the position vector matrix of the i-th dipole.
[0089] (5d) Perform singular value decomposition on the position vector matrix.
[0090] C=U∑V T
[0091] Where U is an m×n orthogonal matrix; V is an m×n orthogonal matrix; and Σ is a diagonal matrix.
[0092] (5f) Solve for the dipole moments of the dipole group in the dataset.
[0093] M=V∑ + U T E mask
[0094] Where, ∑ + It is the pseudo-inverse of matrix Σ in step (5d).
[0095] This ensures that the calculated results of the infinitesimal dipole group match the simulation results, and organizes the dipole moments of different units in the sample set into a dipole moment dataset.
[0096] Step 6, Dataset Processing
[0097] The data in the dataset is first normalized. Since the data in the dataset consists of complex numbers, it is divided into real part datasets and imaginary part datasets for separate training. Then, it is divided into training set, validation set and test set in a ratio of 7:2:1. The curvature features of each sample are used as labels to annotate the dataset.
[0098] Step 7, Building the Long Short-Term Memory Deep Neural Network Model
[0099] Based on the dataset, a long short-term memory deep neural network model is built, including the following steps:
[0100] In the established LSTM network model, the input layer contains two time steps, with the input feature dimension being 1 at each time step, and 50 samples participating in training each time; the hidden layer has two layers, with 256 neurons in each layer; the dropout ratio between LSTM layers is 0.3 to prevent overfitting; and the output layer has 150 neurons, which is the same dimension as the final regression target.
[0101] Step 8: Train and validate the network model.
[0102] (8a) The learning rate during training is set to 0.0001. To learn the features of the data, the learning rate decays to 0.85 every 65 steps. The loss function is MSE, and early stopping logic is used: if the loss value does not decrease within 50 steps, training is stopped to prevent overfitting. The loss function MSE is as follows:
[0103]
[0104] Where S represents the numerical result predicted by the neural network. This represents the actual numerical result.
[0105] (8b) After the model training is completed, the input features in the validation set are input into the neural network to validate the predicted data against the real data.
[0106] Step 9: Use a neural network model to predict the scattering field of the frequency-selective unit.
[0107] (9a) The trained neural network model is used as a database. The predicted data is denormalized and then combined with the infinitesimal dipole position vector matrix C to form a predicted dipole group.
[0108] (9b) Substitute the dipole group into the formula for calculating the scattering pattern of the frequency-selective unit to obtain the scattering field of the frequency-selective surface unit.
[0109] The formula for calculating the scattering pattern of a frequency-selective element is as follows:
[0110]
[0111] Where E is the electric field intensity at the observation point of the scattered field; f(θ,φ) represents the scattering pattern of the frequency-selective unit; (θ,φ) is the observation angle; I is the excitation current; and R is the scattering field distance.
[0112] The advantages of this invention can be further illustrated by the following simulation examples.
[0113] 1. Simulation parameters
[0114] Frequency Selective Unit Information in Figure 2The diagram shows that the operating frequency is 8.56 GHz. The curvature information of the primary and secondary surfaces is given in the form of pairs. Two sets of results are selected for case analysis. The curvatures of these elements are (108, 145) and (600, 90), respectively.
[0115] 2. Simulation Content and Results
[0116] Figure 2 The diagram illustrates how the frequency selection unit in this invention is constructed. Figure 3 A schematic diagram of the infinitesimal dipole model equivalence method in this invention is shown. Figure 4 Two frequency selection units are given in the figure to verify the accuracy of the invention, and the shape information is shown in the figure. Figure 5 and Figure 6 The figures in the middle are a comparison of the predicted and actual real part values for unit 1 and unit 2, respectively. Figure 7 and Figure 8 The predicted values of the imaginary parts of Unit 1 and Unit 2 are compared with the actual values. Table 1 shows the error between the prediction and the actual results of the two units, which proves that the neural network established in this invention has accurate prediction ability.
[0117] Table 1 shows the errors between the predicted and actual results for the two frequency-selective surface elements.
[0118]
[0119] Figure 9 and Figure 10 The paper presents a comparison between the scattering field predicted by the network model and the scattering field simulation results from commercial software. The results are consistent with the analysis results from the commercial software FEKO, verifying the accuracy of the method of the present invention.
[0120] This invention is not limited to the above-described embodiments. Based on the technical solutions disclosed in this invention, those skilled in the art can make some substitutions and modifications to some of the technical features without creative effort, and all such substitutions and modifications are within the protection scope of this invention.
Claims
1. A method for predicting the scattering field of a frequency-selective surface unit with variable curvature, characterized in that, Includes the following steps: Logarithmically sample the curvature of the frequency-selective surface unit and convert it to the original spatial curvature to generate a curvature sampling point sample set. A frequency-selective surface unit model library is established based on the curvature sampling point sample set; The established frequency-selective surface element model library was simulated and solved to obtain near-field data of frequency-selective surface elements with different curvatures. Select the surface unit model library based on the frequency and solve for the infinitesimal dipole position vector matrix; Based on the near-field data of the surface unit and the infinitesimal dipole position vector matrix selected according to different curvature frequencies, the dipole moment in the infinitesimal dipole group is solved to obtain the dipole moment dataset of different units. Based on the dipole moment dataset, a long short-term memory deep neural network model is constructed by using the curvature of the frequency-selected surface unit as input and the dipole moment as output. Based on the long short-term memory deep neural network model, adjust the hyperparameters of the long short-term memory deep neural network model and train the long short-term memory deep neural network model; The dipole moment is predicted based on the trained long short-term memory deep neural network model, and the scattering field of the frequency-selective surface unit is solved by combining the infinitesimal dipole group position vector matrix. The prediction of the scattering field of the frequency-selective surface unit is then completed. Solving for the scattering field of the frequency-selective surface element includes: The trained long short-term memory deep neural network model is used as a database. The predicted data is denormalized and then combined with the infinitesimal dipole position vector matrix to form a predicted dipole group. Substituting the predicted dipole group into the formula for calculating the scattering pattern of the frequency-selective unit, the scattering field of the frequency-selective surface unit is obtained.
2. The method for predicting the scattering field of a variable curvature frequency-selective surface unit according to claim 1, characterized in that, Logarithmic sampling is performed on the curvature of the frequency-selective surface unit. For the arc, sampling is performed on the logarithmic scale. The curvature on the logarithmic scale is restored to the original spatial curvature sampling points to generate a sample set.
3. The method for predicting the scattering field of a variable curvature frequency-selective surface unit according to claim 1, characterized in that, Establish a frequency-selective surface element model library, including: Let the two orthogonal planes XOZ and YOZ be the primary and secondary circular arc surfaces, respectively. The geometry within the primary circular arc surface is a closed surface formed by two circular arcs and two line segments. The secondary circular arc surface is composed of a single circular arc. The sweeping section within the primary circular arc surface is scanned along the circular arc trajectory within the secondary circular arc surface to generate the geometry.
4. The method for predicting the scattering field of a variable curvature frequency-selective surface unit according to claim 1, characterized in that, Simulations were performed on the established frequency-selective surface element model library, including: The established frequency-selective surface element solid model was used in a full-wave simulation software to set observation points, plane wave form, and solution options to solve for near-field data of frequency-selective surface elements with different curvatures.
5. The method for predicting the scattering field of a variable curvature frequency-selective surface unit according to claim 1, characterized in that, Based on the frequency, select a surface element model library and solve for the infinitesimal dipole position vector matrix, including: Discretize the circular arcs in the YOZ plane into nodes to obtain the position information of the dipoles; Rotate the dipole coordinates on the inner arc of the YOZ plane along the X-axis to obtain the dipole position coordinates on the surface unit of the frequency selection surface. Three orthogonally arranged dipoles are placed at each position, and the pointing information and position information are organized into a position vector matrix.
6. The method for predicting the scattering field of a variable curvature frequency-selective surface unit according to claim 1, characterized in that, Solving for the dipole moments in an infinitesimal dipole system yields a dataset of dipole moments for different elements, including: Combining the free space and the Losglin function, based on the infinitesimal dipole model, the expression for the scattering near field is represented. The common factor in the scattering field expression of the infinitesimal dipole model is extracted and rewritten as the product of the dipole moment and the position vector matrix of the infinitesimal dipole. Perform singular value decomposition on the position vector matrix; Solve for the dipole moments of the infinitesimal dipole group so that the field calculation results of the infinitesimal dipole group are in agreement with the full-wave simulation results, and organize the dipole moments of different units in the sample set into a dipole moment dataset.
7. The method for predicting the scattering field of a variable curvature frequency-selective surface unit according to claim 6, characterized in that, The scattering field expression of the infinitesimal dipole model is as follows: in, j For complex units; Angular frequency; Permeability in free space; The first i Angles between an infinitesimal dipole and the X, Y, and Z axes; These are the unit vectors in the X, Y, and Z directions, respectively; The dipole moment of an infinitesimal dipole; i The numbering of the dipole. N d For the total number of dipoles, E The scattering near-field results are obtained using dipoles; It is a dyadic Green's function; Extract the common factor of the scattering field expression of the infinitesimal dipole model and rewrite it as a product of the dipole moment and the position vector matrix: in, E mask The results are from the scattering near-field in the full-wave simulation software. It is the first i Position vector matrix of a dipole; Solve for the dipole moment of an infinitesimal dipole system: in, It is an orthogonal matrix; It is an orthogonal matrix; It is a diagonal matrix. diagonal matrix The pseudo-inverse matrix.
8. The method for predicting the scattering field of a variable curvature frequency-selective surface unit according to claim 1, characterized in that, Building a long short-term memory deep neural network model includes: The Long Short-Term Memory (LSTM) deep neural network model consists of an input layer and hidden layers. The input layer has two time steps, with each time step having a feature dimension of 1, and 50 samples are used for training each time. The hidden layer has two layers, each with 256 neurons. The dropout ratio between LSTM layers is 0.3 to prevent overfitting. The output layer has 150 neurons, which is the same dimension as the final regression target.
9. The method for predicting the scattering field of a variable curvature frequency-selective surface unit according to claim 1, characterized in that, Based on the Long Short-Term Memory (LSTM) deep neural network model, the hyperparameters of the LTM deep neural network model are adjusted, and the LTM deep neural network model is trained, including: The data in the dataset is first normalized. Since the data in the dataset consists of complex numbers, it is divided into real part dataset and imaginary part dataset. Then, it is divided into training set, validation set and test set in a ratio of 7:2:
1. The curvature features of each sample are used as labels to annotate the dataset. The long short-term memory deep neural network model was trained with a learning rate of 0.0001 during training, which was reduced to 0.85 every 65 steps. Training was stopped if the loss value did not decrease within 50 steps using early stopping logic.
10. The method for predicting the scattering field of a variable curvature frequency-selective surface unit according to claim 1, characterized in that, The formula for calculating the scattering pattern of a frequency-selective element is as follows: in, The electric field intensity at the observation point of the unit scattering field; This represents the scattering pattern of the frequency-selective unit. For observation angle; For excitation current; j For complex units; k The propagation coefficient; R The distance is the distance of the scattered field.