Method and system for rapidly predicting electromagnetic scattering of metasurface loading type cavity
By constructing a convolutional residual CNN-ResNet network and combining it with the ADAM optimizer, the problems of high computational cost and insufficient accuracy in electromagnetic scattering prediction of electrically large cavities are solved, and fast and accurate electromagnetic scattering prediction is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-03-17
AI Technical Summary
Existing electromagnetic scattering prediction methods are inefficient for designing electrically large targets such as cavities, have high computational costs, and lack prediction accuracy, especially in complex structures where their learning ability is insufficient.
A convolutional residual CNN-ResNet network is constructed, combined with the ADAM optimizer, and the broadband scattering characteristics of the cavity model are calculated using the full-wave electromagnetic algorithm. The training set and test set are divided, and the network is trained using the backpropagation algorithm to achieve fast electromagnetic scattering prediction.
It reduces the cost of initial data annotation and computation, improves the ability to learn complex features, enhances prediction accuracy, and meets the requirements of engineering practice.
Smart Images
Figure CN121683448A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electromagnetic simulation technology, and specifically relates to a method and system for rapid prediction of electromagnetic scattering, which can be used for rapid prediction of electromagnetic scattering of targets in metasurface-loaded cavity structures. Background Technology
[0002] The electromagnetic stealth performance of an aircraft can be quantitatively characterized by its radar cross section (RCS). Reducing the RCS through proper design can effectively improve its stealth capabilities. Among these, electrically large openings such as engine air intakes, exhaust nozzles, and cockpits are major sources of electromagnetic scattering for aircraft, contributing significantly to the overall RCS.
[0003] Current methods for reducing the RCS of cavities mainly include shape optimization design and the application of absorbing materials. However, these methods all require a large amount of forward modeling calculations and optimization iterations, resulting in low design efficiency. In recent years, metasurface technology has provided a new method for RCS reduction due to its ability to flexibly control the wavefront phase of electromagnetic waves and control scattering through anomalous reflection rather than shape structure. However, these methods also require a large amount of forward modeling calculations and optimization iterations, resulting in low design efficiency. The metasurface control design process is still constrained by the high cost of numerically solving Maxwell's equations, and generally relies on point-by-point optimization using the finite element method (FEM) or the time-difference time-domain (FDTD) algorithm, leading to significant computational bottlenecks and making it difficult to support multi-objective collaborative design. To solve the complex nonlinear mapping problem between the metasurface loading structure and the spectral response, researchers have begun to introduce deep learning methods and apply them to modeling and optimization design, significantly improving prediction accuracy and optimization efficiency.
[0004] Y. Teng, in his paper "Efficient design method for terahertz broadbandmetasurface patterns via..." Deep learning utilizes convolutional neural networks (CNNs) and genetic optimization algorithms (GAs) to achieve rapid prediction of the amplitude and phase response of loaded metasurface structures and efficient optimization of the encoding structure. However, this type of network requires a large amount of simulation datasets, which greatly increases the computational cost and reduces the efficiency of the overall design. Furthermore, this network has insufficient learning ability when dealing with complex metasurface units, resulting in low accuracy of the prediction method after training.
[0005] JH Fu in the paper "Highly-efficient design method for coding metasurfacesbased on deep The deep neural network (DNN) cascade paradigm proposed by “learning” significantly accelerates the optimization process of inferring structural encoding from the target spectrum through the collaboration of forward and backward networks. However, its construction of datasets still requires a large amount of simulation computation. Similarly, its disadvantages are high computational cost and long design cycle.
[0006] Patent application number 2024104746255 discloses an automatic prediction method for the response characteristics of digitally encoded metasurfaces based on the Finite-Difference Time-Domain (FDTD) method and convolutional neural networks (CNNs). This method constructs a dataset using the FDTD algorithm and utilizes a CNN architecture to predict the electromagnetic response of metasurfaces with arbitrary geometric features. While this method and system are suitable for predicting the response characteristics of simple, small-structure metasurfaces, for complex, electrically large structures such as cavity targets, the computational load of electromagnetic scattering increases significantly, leading to a substantial increase in computational time and cost.
[0007] In summary, existing methods for electromagnetic scattering prediction suffer from two main problems: First, they all rely on large-scale, high-quality labeled data, which significantly increases upfront computational costs. Specifically, when used for electromagnetic control design of electrically large targets such as cavity-loaded metasurfaces, the surge in electromagnetic nonlinearity makes data generation far more time-consuming than training. Second, the training network architecture lacks the ability to learn from complex models, resulting in prediction accuracy that fails to meet engineering requirements. Therefore, a robust and efficient electromagnetic scattering prediction design method is urgently needed. Summary of the Invention
[0008] The purpose of this invention is to overcome the shortcomings of the existing technology and propose a fast prediction method and system for electromagnetic scattering of metasurface-loaded cavities, so as to effectively reduce the cost of early data annotation and calculation, improve the learning ability of complex features, and improve prediction accuracy.
[0009] To achieve the above objectives, the technical solution of the present invention includes the following:
[0010] 1. A method for rapid prediction of electromagnetic scattering in a metasurface-loaded cavity, characterized in that it comprises:
[0011] (1) Establish a cuboid cavity model, and load metasurface array units onto the inner and outer surfaces of its three selected faces to form a metasurface-loaded cavity model;
[0012] (2) Binary encoding is performed based on the phase of each basic metasurface unit to construct an encoding matrix containing phase and position information of the metasurface array surface. ;
[0013] (3) Construct a convolutional residual CNN-ResNet network that includes a convolutional neural network (CNN) and a residual network (ResNet);
[0014] (4) The wideband scattering characteristics RCS of the cavity model are calculated based on the full-wave electromagnetic algorithm, and the training set and test set are divided.
[0015] (5) Input the training set into the CNN-ResNet network model, use the ADAM optimizer to iteratively train it, continuously update and optimize the model parameters until the set error threshold is met, and obtain the trained convolutional residual CNN-ResNet network model.
[0016] (6) Input the test set into the trained convolutional residual CNN-ResNet network model to obtain the fast prediction results of electromagnetic scattering of digitally encoded metasurface loaded cavity.
[0017] Furthermore, the construction of the convolutional residual CNN-ResNet network includes the following implementation:
[0018] 3a) Construct a convolutional neural network (CNN) consisting of convolutional layers, batch normalization layers, flattening layers, and fully connected layers connected in sequence;
[0019] 3b) Construct a ResNet layer, which consists of two concatenated convolutional layers followed by a residual layer and a fully connected layer.
[0020] 3c) Connect the two parts of the above-constructed convolutional neural network (CNN) layer and residual network (ResNet) layer in sequence to form a CNN-ResNet network.
[0021] 2. A rapid prediction system for electromagnetic scattering of a metasurface-loaded cavity, characterized in that it comprises:
[0022] The metasurface encoding matrix generation module is used to generate a matrix consisting of digits 0 and 1 that contains phase information and position feature information of metasurface units;
[0023] The full-wave electromagnetic calculation model construction module is used to construct the simulation model of the corresponding metasurface based on the 0 / 1 matrix, and to calculate its broadband scattering characteristic RCS using the full-wave electromagnetic algorithm.
[0024] The dataset construction module is used to set different frequency bands and simulation parameters, perform simulation calculations on the full-wave electromagnetic model, and collect RCS data of the electromagnetic scattering characteristics of the metasurface. Using the RCS data and 0 / 1 matrix, the dataset is created and divided into training and test sets.
[0025] The network model building module is used to construct a CNN-ResNet network consisting of convolutional neural network sub-modules and residual network unit sub-modules.
[0026] The network model training module is used to calculate the root mean square error (RMSE) of the loss function using the training set, and to calculate its gradient using the backpropagation algorithm. The gradient descent method is used to train the CNN-ResNet network.
[0027] The prediction module is used to add the test set to the trained CNN-ResNet network model to quickly predict the RCS of the electromagnetic scattering properties of the digitally encoded metasurface.
[0028] 3. An electronic device, characterized in that it comprises:
[0029] Processor and memory;
[0030] The memory stores program instructions;
[0031] The processor is used to run the program instructions to execute the rapid prediction method for electromagnetic scattering of the metasurface-loaded cavity.
[0032] 4. A computer-readable storage medium, characterized in that the computer-readable storage medium stores program instructions, which are executed to perform the rapid prediction method for electromagnetic scattering of the metasurface-loaded cavity.
[0033] This invention constructs a convolutional residual CNN-ResNet network model and combines iterative training with the ADAM optimization algorithm. Compared with traditional full-wave electromagnetic algorithms, it overcomes the shortcomings of relying on fine mesh partitioning and iterative solution, resulting in high computational cost and difficulty in meeting the engineering requirements of real-time design and optimization. Compared with existing deep learning methods such as CNN networks, its trained network has better prediction performance and smaller error, and can realize the prediction of electromagnetic scattering characteristics of loaded metasurface cavity targets, thus improving prediction accuracy. Attached Figure Description
[0034] Figure 1 This is a flowchart illustrating the rapid prediction method for electromagnetic scattering of the metasurface-loaded cavity of the present invention.
[0035] Figure 2 This is a cavity model diagram of the loading metasurface array set in the method of the present invention;
[0036] Figure 3 This is a diagram showing the correspondence between the phase and matrix of the metasurface unit in the method of this invention;
[0037] Figure 4 This is a schematic diagram of the CNN-ResNet network structure constructed in the method of the present invention;
[0038] Figure 5 This is a block diagram of the electromagnetic scattering rapid prediction system for the metasurface-loaded cavity of the present invention.
[0039] Figure 6 This is a diagram of the electronic device of the rapid electromagnetic scattering prediction system for the metasurface-loaded cavity of the present invention;
[0040] Figure 7 The image shows a comparison of the prediction results of electromagnetic scattering characteristics RCS using the method of this invention and the existing CNN convolutional network method, respectively. Detailed Implementation
[0041] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0042] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.
[0043] Example 1: A fast prediction method for electromagnetic scattering of metasurface-loaded cavities.
[0044] Reference Figure 1 The implementation steps of this example include the following:
[0045] Step 1: Construct a metasurface-loaded cavity model.
[0046] Reference Figure 2 The implementation of this step includes:
[0047] 1.1) Establish a cuboid cavity model, such as... Figure 2 As shown in (a), three faces are selected on the model;
[0048] 1.2) Load metasurface array elements onto the inner and outer surfaces of three selected faces:
[0049] 1.2.1) Based on the cavity model dimensions and the structural parameters of the metasurface basic units, determine the array spacing and arrange the metasurface basic units in a periodic array to form a metasurface array unit, such as... Figure 2 As shown in (b);
[0050] 1.2.2) On the opening face and the other two sides of the selected cavity, arrange the metasurface array elements at fixed intervals on the inner and outer surfaces of the selected cavity to form a metasurface-loaded cavity electromagnetic model, such as... Figure 2 As shown in (c).
[0051] Step 2: Generate the metasurface 0 / 1 encoding matrix.
[0052] Reference Figure 3 The implementation of this step includes:
[0053] 2.1) The encoding matrix consists of digits 0 and 1. It contains phase information and spatial position information of the metasurface array units;
[0054] 2.2) Determine the matrix row and column numbers based on spatial location information. The row number represents the horizontal position information of the metasurface unit, and the column number represents the vertical position information of the metasurface unit. Based on the phase encoding information, assume that the encoded value 0 corresponds to a unit phase of 0°, and the encoded value 1 corresponds to a unit phase of 180°. Determine the encoding matrix through the encoded phase information and spatial information. It is represented as follows:
[0055] ,
[0056] in, and Representing the global matrix respectively The number of rows and columns in the text, here , 1 indicates a unit phase of 180°, and 0 indicates a unit phase of 0°.
[0057] Step 3: Construct the CNN-ResNet network.
[0058] Reference Figure 4 The implementation of this step includes:
[0059] 3.1) Establish the layers of a Convolutional Neural Network (CNN):
[0060] 3.1.1) A first convolutional layer with a 2×2 three-channel convolutional kernel is selected to extract features from the input 0 / 1 encoding matrix;
[0061] 3.1.2) Set up a batch normalization layer to standardize data through max pooling operations;
[0062] 3.1.3) Set up a flattening layer to reduce the dimensionality of the feature data to a one-dimensional vector;
[0063] 3.1.4) Set up a first fully connected layer with multiple weight parameters to initially learn global feature relationships. Its output is used as input to the ResNet network to further refine the deep feature representation.
[0064] 3.1.5) The convolutional layer, batch normalization layer, flattening layer and fully connected layer are connected sequentially to form the CNN network layers;
[0065] 3.2) Establish the ResNet residual network layer:
[0066] 3.2.1) A second convolutional layer consisting of two 3×3 four-channel convolutional kernels connected in series is selected to extract features from the output data of the CNN layer;
[0067] 3.2.2) The output of the second convolutional layer is stacked element by element with the input to form a residual layer, which is used to capture the relationship between changes in local features and the overall output;
[0068] 3.2.3) Set up a second fully connected layer with multiple weight parameters to output the mapping results of the residual layer;
[0069] 3.2.4) The second convolutional layer, the residual layer, and the second fully connected layer are sequentially connected to form a ResNet layer.
[0070] 3.3) Connect the two parts of the above-convolutional neural network (CNN) layer and residual network (ResNet) layer in sequence to form a convolutional residual CNN-ResNet network.
[0071] Step 4: Simulate and construct a dataset for the metasurface-loaded cavity model.
[0072] 4.1) Based on the metasurface encoding matrix The model includes phase and spatial position information of the metasurface array elements. It is used to design a metasurface-loaded cavity model, set the simulation frequency band, and determine the simulation calculation angle.
[0073] 4.2) The RCS of the scattering characteristics of the multiple models in the above design is calculated using the full-wave electromagnetic algorithm to form multiple model datasets. Each dataset includes the RCS and its corresponding encoding matrix. The datasets are divided into training set and test set in an 8:2 ratio.
[0074] Step 5: Train the convolutional residual CNN-ResNet network.
[0075] 5.1) Set the maximum number of network iterations to K and set the root mean square error (RMSE) threshold T;
[0076] 5.2) Input the encoding matrix from the training set into the CNN-ResNet network to output the predicted value, and use this predicted value to calculate the root mean square error (RMSE) of the loss function:
[0077] ,
[0078] in, This represents the true RCS value calculated by the model corresponding to the encoding matrix in the training set at the i-th frequency point.
[0079] This represents the RCS prediction value of the CNN-ResNet network output from the encoding matrix in the training set at the i-th frequency point. This represents the number of frequency sampling points.
[0080] 5.3) The RMSE gradient of the CNN-ResNet network is calculated using the backpropagation algorithm;
[0081] 5.4) The ADAM optimizer is used to adaptively adjust the first-order moment estimates and second-order moment estimates, combined with a dynamic learning rate mechanism, to update the network weight parameters;
[0082] 5.5) Repeat steps (5.2)-(5.4) above until the RMSE is lower than the preset threshold or the maximum number of iterations K is reached, then terminate the training to obtain the trained convolutional residual CNN-ResNet network model.
[0083] Step 6: Output the RCS prediction results of the scattering characteristics.
[0084] The encoding matrix in the test set is input into the trained CNN-ResNet network model to quickly predict the RCS of the electromagnetic scattering properties of the digitally encoded metasurface, thus obtaining the RCS prediction results of the surface-loaded cavity model.
[0085] It should be noted that the flowchart representations or method representations of the above embodiments can be understood as representing a module, segment, or portion of code comprising one or more executable instructions configured to implement a specific logical function or process. This invention is not limited to the disclosed preferred embodiments, and its implementation may not follow the order shown or discussed. That is, the step numbers in the specification and claims are only for clear description and understanding of the embodiments of this invention, and their order is not limited.
[0086] Example 2: Rapid prediction system for electromagnetic scattering of metasurface-loaded cavities.
[0087] Reference Figure 5 This example includes: a metasurface encoding matrix generation module 1, a full-wave electromagnetic computation model construction module 2, a dataset construction module 3, a network model construction module 4, a network model training module 5, and a prediction module 6. The network model construction module 4 includes a convolutional neural network (CNN) submodule 41, a residual network (ResNet) unit submodule 42, and a splicing submodule 43. The CNN submodule 41 includes a first convolutional layer 411, a batch normalization layer 412, a flattening layer 413, and a first fully connected layer 414. The ResNet submodule 42 includes a second convolutional layer 421, a residual layer 422, and a second fully connected layer 423.
[0088] The working principle of the entire system is as follows:
[0089] The metasurface encoding matrix generation module 1 is used to generate a matrix composed of digital 0 / 1 that contains phase information and position feature information of metasurface units, and transmit the encoding matrix information to the full-wave electromagnetic calculation model construction module 2;
[0090] The full-wave electromagnetic calculation model construction module 2 is used to construct the corresponding metasurface simulation model based on the encoding matrix obtained in the metasurface encoding matrix generation module 1, set different frequency bands and simulation parameters, simulate the full-wave electromagnetic model and calculate the broadband scattering characteristics RCS, and then transmit the calculated RCS values of multiple sets of metasurface loaded cavity models to the dataset construction module 3.
[0091] The dataset construction module 3 is used to construct multiple model datasets from the RCS values of multiple sets of metasurface loaded cavity models obtained in the full-wave electromagnetic calculation model construction module 2. Each dataset includes the RCS and its corresponding encoding matrix. The dataset is divided into a training set and a test set in an 8:2 ratio and then transmitted to the network model training module 5 and the prediction module 6, respectively.
[0092] The network model building module 4 is used to build a convolutional residual CNN-ResNet network, wherein:
[0093] The Convolutional Neural Network (CNN) submodule 41 is used to establish a CNN network layer including a first convolutional layer 411, a batch normalization layer 412, a flattening layer 413, and a first fully connected layer 414, for preliminary feature extraction of input data.
[0094] The residual ResNet network unit submodule 42 is used to establish a residual ResNet network layer including a second convolutional layer 421, a residual layer 422, and a second fully connected layer 423, for extracting local features from the input data.
[0095] The splicing submodule 43 is used to connect the convolutional neural CNN network submodule 41 and the residual ResNet network unit submodule 42 to form a convolutional residual CNN-ResNet network;
[0096] The network model training module 5 is used to calculate the root mean square error (RMSE) of the loss function using the training set, calculate its gradient using the backpropagation algorithm, train the CNN-ResNet network using the gradient descent method, and transmit the trained CNN-ResNet network model to the prediction module 6.
[0097] Prediction module 6 is used to quickly predict the RCS of the digitally encoded metasurface electromagnetic scattering properties of the test set in the trained CNN-ResNet network model and output the RCS curve.
[0098] It should be noted that the above functional modules can be implemented, in whole or in part, through software, hardware, or any combination thereof. When implemented in software, they can be implemented, in whole or in part, as program instruction products. A program instruction product includes one or a set of program instructions. When the program instructions are loaded and executed on a computer, the described process or function is generated, in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Program instructions can be stored in a computer-readable and writable storage medium, or transferred from one computer's readable and writable storage medium to another.
[0099] The modules shown or discussed in this embodiment can be implemented through direct module coupling. The various functional modules and sub-modules in this embodiment can dynamically reside within a single processing unit, or each module can exist physically independently, or two or more modules can dynamically reside within a single processing unit. When these dynamic components are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable and writable storage medium. This storage medium can be a memory, disk, or optical disc, etc.
[0100] Example 3, Electronic Device
[0101] Reference Figure 6 This example includes: a processor S1, a memory S2, a network input / output interface S3, and a bus S4. The processor S1, memory S2, network input / output interface S3, and bus S4 enable communication connections between them within the device.
[0102] The processor S1 can be implemented using a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to construct the CNN-ResNet network of the present invention, so as to realize the technical solution provided in the embodiments of this specification.
[0103] The memory S2 can be implemented in the form of read-only memory (ROM), random access memory (RAM), static storage device, dynamic storage device, etc., to store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory S2 and is called by the processor S1 to execute the steps in the rapid prediction method for electromagnetic scattering of the metasurface loaded cavity described in any one of the embodiments.
[0104] The network input / output interface S3 is used to connect to the dataset construction / prediction module to realize information input and output. This network input / output / module can be configured as a component within the device or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touchscreens, microphones, various sensors, etc., while output devices may include displays, speakers, vibrators, indicator lights, etc.
[0105] The bus S4 includes a path for transmitting information between various components of the device, such as the processor S1, memory S2, and network input / output interface S3.
[0106] It should be noted that although the above device only shows the processor S1, memory S2, network input / output interface S3, and bus S4, in actual implementation, the device may also include other components necessary for normal operation.
[0107] Example 4, readable storage medium.
[0108] This invention provides a computer-readable storage medium storing multiple instructions that can be loaded by a processor to execute any step in the rapid prediction method for electromagnetic dispersion of metasurface-loaded cavities provided in this invention.
[0109] It should be noted that the computer-readable storage medium in this embodiment includes both permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), erasable and programmable read-only memory (EPROM), flash memory or other memory technologies, optical discs, CD-ROMs, digital versatile optical discs (DVDs), or other optical storage, magnetic tape, magnetic magnetic disks, magneto-optical discs, or other magnetic storage devices or any other non-transfer medium, which can be used to store information accessible by a computing device.
[0110] The technical effects of the present invention will be further explained below with reference to simulation results:
[0111] I. Simulation conditions.
[0112] Simulation software: Matlab 2024b + Python 4.12 + PyCharm 2024.1 + Feko;
[0113] Simulation parameters: The designed cavity dimensions are 160×165.5×171mm.3 The dimensions of a single metasurface basic unit are length L = 8.1 mm, W = 8.1 mm, and the length of the C-shaped opening B = 1.8 mm. A single metasurface array unit consists of 16*16 basic units. The simulation frequency band is 6-16 GHz, and the simulation angle is... .
[0114] II. Simulation Content.
[0115] Using the simulation software described above, the RCS (Radio Reflectance Characteristic) of the metasurface-loaded cavity model designed with the above simulation parameters was predicted using both the method of this invention and an existing automatic prediction method for digitally coded metasurface response characteristics based on the finite-difference time-domain method and convolutional neural networks. Furthermore, the RCS of the metasurface-loaded cavity model was calculated using a full-wave electromagnetic algorithm. The results are as follows: Figure 7 The triangular-marked curve represents the RCS result predicted by the CNN-ResNet network in the method of this invention, the circular-marked curve represents the RCS result predicted by the CNN network in existing methods, and the unmarked curve represents the RCS result calculated by the full-wave electromagnetic algorithm.
[0116] from Figure 7 As can be seen from the comparison data of the curves, the RCS curve of the method of the present invention is closer to the full-wave electromagnetic calculation result than the RCS curve of the existing method, indicating that the prediction result of the present invention has higher accuracy.
[0117] By calculating the RMSE values of the curve results using the formula, the RMSE values of the prediction results obtained by the present invention and the existing methods are 0.2255 dB and 0.4884 dB, respectively. The RMSE value of the present invention is smaller than that obtained by the existing methods, further demonstrating that the prediction results obtained by the present invention through the construction of a CNN-ResNet network model have higher accuracy.
[0118] It should be noted that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; under the concept of the present invention, the technical features of the above embodiments or different embodiments can also be combined, and there are many other variations of the different aspects of the present invention as described above, which are not provided in detail for the sake of brevity.
[0119] Furthermore, to simplify the description and discussion, and to ensure the invention remains understandable, the process of optimizing the metasurface basic unit model may or may not be shown in the provided drawings. While specific details have been set forth to describe exemplary embodiments of the invention, it will be apparent to those skilled in the art that the invention may be practiced without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.
Claims
1. A fast prediction method of electromagnetic scattering of a super surface loaded cavity, characterized in that, The method comprises the following steps: (1) a cuboid cavity model is established, and a metasurface array unit is loaded on the selected three faces of the cuboid cavity model to form a metasurface-loaded cavity model; (2) Binary coding based on the phase of each metasurface basic unit, constructing an encoding matrix containing the phase information and position information of the metasurface array surface ; (3) a convolution residual CNN-ResNet network is constructed, which comprises a convolution neural network CNN and a residual network ResNet; (4) the wideband scattering characteristics RCS of the cavity model are calculated based on a full-wave electromagnetic algorithm, and a training set and a test set are divided; (5) the training set is input into the CNN-ResNet network model, and an ADAM optimizer is used for iterative training, and the model parameters are updated until the error threshold is met, and a trained convolution residual CNN-ResNet network model is obtained; (6) the test set is input into the trained convolution residual CNN-ResNet network model, and the electromagnetic scattering fast prediction result of the digital coded metasurface-loaded cavity is obtained.
2. The method of claim 1, wherein, In step (1), the metasurface array unit is loaded on the selected three faces of the cuboid cavity model to form a metasurface-loaded cavity model, and the implementation comprises: 1a) according to the size of the cavity model and the structural parameters of the metasurface basic unit, the array spacing is determined to arrange the metasurface basic unit in a periodic array to form a metasurface array unit; 1b) select the opening corresponding face and the other two side faces of the cavity, and arrange the metasurface array unit on the selected inner and outer surfaces of the cavity at a fixed spacing to form a metasurface-loaded cavity electromagnetic model.
3. The method of claim 1, wherein, constructing an encoding matrix comprising metasurface array face phase information and position information in step (2) implementations include: 2a) the binary code (0 / 1) is used to define the discrete phase distribution of the metasurface basic unit: the code value 0 corresponds to the unit phase 0°, and the code value 1 corresponds to the unit phase 180°; 2b) Determining the matrix element values from the encoded phase information, determining the matrix row and column numbers from the spatial position information, determining the encoding matrix from the encoded phase information and the spatial information is represented as follows: ， wherein and represent the number of row and column cells in the global matrix respectively, here , 1 means a cell phase of 180° and 0 means a cell phase of 0°.
4. The method of claim 1, wherein, In step (3), the convolution residual CNN-ResNet network is constructed, and the implementation comprises: 3a) a convolution neural network CNN layer is established, which comprises a convolution layer, a batch normalization layer, a flattening layer and a fully connected layer connected in sequence, wherein: the convolution layer uses a 2x2 three-channel convolution kernel to extract features from the input 0 / 1 code matrix; the batch normalization layer realizes data standardization through maximum pooling operation; the flattening layer is used to reduce the feature data to a one-dimensional vector; the fully connected layer is used to preliminarily learn the global feature relationship, and the output thereof is used as the input of the ResNet network to further optimize the deep feature expression; 3b) a residual network ResNet layer is established, which comprises two convolution layers connected in series and then connected with a residual layer and a fully connected layer, wherein: each convolution layer comprises a 3x3 four-channel convolution kernel for extracting features from the CNN layer output data; the residual layer is used to add the output and the input of the convolution layer element by element to capture the relationship between the local feature change and the overall output; the fully connected layer is used to map the output of the residual layer to the final output of the CNN-ResNet network model; 3c) the convolution neural network CNN layer and the residual network ResNet layer are connected in sequence to form the CNN-ResNet network.
5. The method of claim 1, wherein, In step (4), the wideband scattering characteristics RCS of the cavity model are calculated based on a full-wave electromagnetic algorithm, and the implementation comprises: 4a) randomly generate multiple sets of encoding matrices composed of digital 0 / 1 According to the metasurface encoding matrix The spatial position information of the metasurface array unit contained in the metasurface is designed, the metasurface loading cavity model is set, the simulation frequency band and the simulation calculation angle are set. 4b) using the full-wave electromagnetic algorithm to calculate the scattering characteristics RCS of the designed multiple groups of models, to form a data set of multiple groups of models corresponding to the encoding matrix information and RCS, and to divide the data set into a training set and a test set in a ratio of 8:
2.
6. The method of claim 1, wherein, In step (5), the CNN-ResNet network model is iteratively trained using the ADAM optimizer, which includes the following steps: 5a) setting the maximum number of network iterations K and setting the root mean square error RMSE threshold T to 3dB; 5b) inputting the training set encoding matrix into the CNN-ResNet network to output a predicted value, and calculating the loss function root mean square error RMSE value using the predicted value: ， wherein, represents the RCS real value calculated by the encoding matrix corresponding model in the training set at the ith frequency point, represents the RCS predicted value output by the encoding matrix CNN-ResNet network in the training set at the ith frequency point, is the number of frequency sampling points; 5c) calculating the RMSE gradient of the CNN-ResNet network using the back propagation algorithm; 5d) using the ADAM optimizer to adaptively adjust the first moment estimate and the second moment estimate, and combining a dynamic learning rate mechanism to update the network weight parameters; 5e) repeating steps (5b)-(5d) until the RMSE is lower than the preset threshold or the maximum number of iterations K is reached, and terminating the training to obtain a trained convolutional residual CNN-ResNet network model.
7. A system for fast prediction of electromagnetic scattering of a super surface loaded cavity, characterized in that, It includes: A metasurface encoding matrix generation module for generating a matrix composed of digital 0 / 1 containing phase information and position characteristic information of metasurface units; A full-wave electromagnetic calculation model construction module for constructing a simulation model of the corresponding metasurface according to the 0 / 1 matrix, setting different frequency bands and simulation parameters, simulating the full-wave electromagnetic model and calculating the wideband scattering characteristics RCS; A data set construction module for collecting electromagnetic scattering characteristics RCS data of the metasurface, using the RCS data and the 0 / 1 matrix to make a data set and divide it into a training set and a test set; A network model construction module for constructing a CNN-ResNet network composed of a convolutional neural network submodule and a residual network unit submodule; A network model training module for calculating the loss function root mean square error RMSE using the training set, and calculating its gradient using the back propagation algorithm, and training the CNN-ResNet network using the gradient descent method; A prediction module for adding the test set to the trained CNN-ResNet network model to quickly predict the electromagnetic scattering characteristics RCS of the digitally encoded metasurface.
8. The system of claim 7, wherein, The network model construction module includes: A CNN network submodule for establishing a CNN network including a convolutional layer, a batch normalization layer, a flattening layer, and a fully connected layer: A ResNet network submodule for constructing a ResNet network including two convolutional layers, a residual layer, and a fully connected layer; A splicing submodule for sequentially connecting the convolutional neural network CNN submodule and the residual network ResNet submodule to form a CNN-ResNet network module.
9. An electronic device, comprising: It includes: A processor and a memory; The memory stores program instructions; The processor is configured to run the program instructions to perform the electromagnetic scattering fast prediction method for a metasurface loaded cavity as claimed in any one of claims 1 to 6.
10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores program instructions, and the program instructions are executed to perform the electromagnetic scattering fast prediction method of the super surface loaded cavity.