A metasurface modulation device and method based on a physical constraint generation model

Abstract

The application discloses a kind of super surface regulation and control device and method based on physical constraint generation model, it is related to electromagnetic super surface scattering regulation and control technical field.Method includes obtaining super surface array and forms different coding array structure, i.e.RCS response;Obtain simulation dataset and calculation dataset, establish including prediction model, identification model and generation model physical constraint depth condition generation model, utilize calculation dataset and simulation dataset to construct formula-simulation residual error network expansion dataset, pre-training is carried out to prediction model, and simulation dataset is used to the joint training of prediction model, identification model and generation model;Input target RCS condition, output optimal super surface unit arrangement scheme, realize the quick simulation and reconstruction of RCS.The application can reduce electromagnetic simulation workload, solve the problem that one-to-many mapping in traditional neural network is difficult to accurately train, and improve the intelligence and stability of super surface simulation RCS in combination with mechanical regulation.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic metasurface scattering control technology, and in particular to a metasurface control device and method based on a physical constraint generation model. Background Technology

[0002] With the rapid development of electromagnetic metamaterials and artificial metasurfaces, the manipulation of radar cross section (RCS) based on metasurfaces has attracted widespread attention in fields such as target stealth, radar signal management, and electromagnetic compatibility. By designing periodic or aperiodic unit structures at the subwavelength scale, metasurfaces can precisely control the scattering characteristics of incident electromagnetic waves, thereby achieving the redistribution of scattered energy and effective control of RCS characteristics.

[0003] Traditional methods for determining the metasurface array arrangement corresponding to a specific RCS typically rely on extensive electromagnetic simulations and complex optimization processes, resulting in high computational costs and long processing times. In recent years, deep learning, as an efficient data-driven approach, has been introduced into the field of metasurface design to handle complex structure-response relationships. Forward prediction models can predict the RCS distribution based on given structural parameters, achieving a one-to-one mapping from structure to performance. Although the RCS distribution of a specific arrangement of metasurfaces can be calculated using the far-field superposition formula, the strong coupling effect between some elements makes it difficult to accurately describe analytically. Therefore, traditional models still rely on electromagnetic simulation software (such as CST and HFSS) to generate a large number of samples, leading to high simulation costs and low efficiency.

[0004] Unlike forward optimization, inverse design aims to generate corresponding metasurface arrangements given a target RCS distribution. Because there is a non-unique mapping between input and output, multiple structures may satisfy the same design requirement, significantly increasing the difficulty of training neural networks. The complex structure-response coupling also causes traditional optimization algorithms (such as particle swarm optimization and genetic algorithms) to easily get trapped in local optima in high-dimensional spaces, making it difficult to obtain the global optimum. Furthermore, each inverse search requires iterative optimization, hindering fast response.

[0005] Furthermore, in existing methods, the forward prediction model and the reverse generation model are mostly trained independently or simply connected in series, lacking necessary physical constraints. This can easily lead to overfitting or physical uninterpretation, thus limiting their application in practical electromagnetic design and engineering implementation.

[0006] On the other hand, mechanically adjustable metasurfaces show great potential for dynamic RCS control. By switching the angle or state of metasurface units through a mechanical control system, real-time structural reconstruction can be achieved at the physical level. However, there is currently a lack of a unified design framework that can organically combine deep generative models with mechanical control hardware systems to achieve fully automated design and control from target RCS input to mechanical structure output.

[0007] Therefore, proposing a metasurface control device and method based on a physical constraint generation model to solve the problems existing in the prior art is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] In view of this, the present invention provides a metasurface control device and method based on a physical constraint generation model, which can reduce the workload of electromagnetic simulation, solve the problem of difficult accurate training of one-to-many mapping in traditional neural networks, and improve the intelligence and stability of metasurface simulation RCS by combining mechanical control.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A metasurface control device based on a physical constraint generation model includes a metasurface array and a mechanical control device. The metasurface array consists of N×N metasurface units, each of which includes a bottom square metal layer, a middle cylindrical dielectric layer, and a top metal pattern layer. The mechanical control device includes a control module, a drive circuit, and stepper motors, with each stepper motor connected to a metasurface unit.

[0010] Optionally, the top metal pattern layer includes two circular I-shaped structures with the same opening angle but different orientation angles, so that the metasurface unit forms two encoding states of "0" and "1". Stepper motors are used to drive metasurface units to switch between "0" and "1" states.

[0011] A metasurface manipulation method based on a physical constraint generation model, applied to a metasurface manipulation device based on a physical constraint generation model as described above, includes the following steps: S1. Obtain the metasurface array and form different encoding array structures according to the encoding state of the metasurface units, corresponding to different RCS responses; S2. Simulation datasets are generated by obtaining simulation RCS data of each coding array through electromagnetic simulation software CST. Simultaneously, calculation datasets are generated by combining theoretical RCS data calculated based on remote superposition formula. Formula-simulation residual network is constructed using calculation datasets and simulation datasets, and calculation datasets and simulation datasets are expanded respectively. S3. Establish a physical constraint depth condition generation model, which includes three parts: a prediction model, an identification model, and a generation model. S4. Use the expanded computational dataset to pre-train the prediction model, establish and quickly fit the physical constraint relationship to form a pre-trained prediction model. Then, use the expanded simulation dataset to jointly train the pre-trained prediction model, recognition model and generation model to obtain the trained model, and realize the parameter co-optimization and response consistency physical constraints between models. S5. Input the conditions of the target RCS into the trained model, output the optimal metasurface unit arrangement scheme, and the mechanical control device controls the metasurface units according to the optimal metasurface unit arrangement scheme to realize the rapid simulation and reconstruction of RCS.

[0012] Optionally, the simulation dataset uses the electromagnetic simulation software CST to simulate the full-space scattering characteristics of different metasurface coded arrays to obtain radar cross section (RCS) data within a predetermined frequency band and angle range. The dataset is computed based on the pattern function of the metasurface to calculate the far-field scattering response with arbitrary phase distribution. The far-field superposition formula is used to theoretically calculate the corresponding theoretical RCS distribution for each coded array. The pattern function of the coded metasurface is used as the basis for this calculation. as follows: , The metasurface is an N×N array, and the size of each unit is... , The abrupt phase change provided by the cell located in the m-th row and n-th column for the j-th operating frequency point is: and These are pitch and azimuth angles. It is the wavenumber at the operating frequency.

[0013] Optionally, the prediction model is based on a convolutional neural network structure. A formula-simulation residual network is constructed to expand the dataset using the computational and simulation datasets. The residual network structure is a one-dimensional convolutional neural network, used to obtain simulation-like calibration data through low-cost formula data extrapolation. This data is then used for pre-training the prediction model. The root mean square error (RMSE) between the predicted and actual RCS values ​​is used as the loss function for optimization. The loss function employs reconstruction loss. The expression is: , in, This is the true RCS value. To predict the RCS value, N is the number of sample points.

[0014] Optionally, the total loss value for joint training is composed of the reconstruction loss. KL divergence regularization Coding reconstruction error With cycle consistency The composition, expressed as: , Among them, reconstruction loss It is the error between the RCS predicted by the prediction model and the true RCS, KL divergence regularization It is the KL divergence between the latent variable distribution and the standard normal distribution, and the encoding reconstruction error. It is the binary cross-entropy between the generated model's output encoding and the actual encoding, and cycle consistency. The consistency between the RCS predicted by the generative model and the true RCS after the generative model generates coded input and predicts the model gives the model physical constraints. , , and These are the weighting coefficients for the four losses.

[0015] Optional, the optimal metasurface unit arrangement schemes include: The generative model samples the target RCS distribution and outputs multiple candidate metasurface coding schemes. The candidate codes are input into the prediction model to obtain the corresponding RCS prediction results; the mixed similarity index between each prediction result and the target RCS is calculated. The mixed similarity includes the root mean square error (RMSE) and the structural similarity of the center angle range. The candidate scheme with the highest hybrid similarity is selected as the optimal metasurface encoding arrangement scheme.

[0016] As can be seen from the above technical solution, compared with the prior art, the present invention provides a metasurface control device and method based on a physical constraint generation model, which has the following beneficial effects: This invention pre-trains the prediction model using a computational dataset and then fine-tunes it using a simulation dataset, reducing data acquisition costs while improving model training efficiency and decreasing the workload of electromagnetic simulation. The deep conditional generative model, composed of the recognition and generative models, can generate multiple candidate metasurface array arrangement schemes under the target RCS condition, solving the convergence problem of one-to-many mapping in traditional reverse engineering. The joint training of the prediction, recognition, and generative models introduces physical constraints during the learning process, improving model stability and physical interpretability. Combined with a mechanical control system, the "0 / 1" state switching of metasurface units can be realized, thereby rapidly reconstructing the target RCS distribution at the physical level and enhancing the intelligence and reconfigurability of metasurface RCS simulation. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 These are diagrams of two metasurface unit structures disclosed in this invention; Figure 2 This is a flowchart of the metasurface control method based on a physical constraint generation model disclosed in this invention; Figure 3 This is a schematic diagram of the metasurface control method based on a physical constraint generation model disclosed in this invention. Figure 4 This is a diagram of the network structure of the prediction model disclosed in this invention; Figure 5 This is a network structure diagram of the recognition model and generation model disclosed in this invention; Figure 6 This is a comparison chart of the simulation results corresponding to the target RCS, model predicted RCS, and model output encoding in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Reference Figure 1 As shown, this invention discloses a metasurface control device based on a physical constraint generation model, comprising a metasurface array and a mechanical control device. The metasurface array consists of N×N metasurface units, each of which includes a bottom square metal layer, a middle cylindrical dielectric layer, and a top metal pattern layer. The mechanical control device includes a control module, a drive circuit, and stepper motors, with each stepper motor connected to a metasurface unit.

[0021] Furthermore, refer to Figure 1 As shown, the top metal pattern layer includes two circular I-shaped structures with the same opening angle but different orientation angles, which makes the metasurface unit form two encoding states of "0" and "1". The two metasurface units have basically the same reflection amplitude in the 8-9 GHz operating frequency band, and the reflection phases differ by about 180°. Stepper motors are used to drive metasurface units to switch between "0" and "1" states.

[0022] A metasurface control method based on a physical constraint generation model is applied to a metasurface control device based on a physical constraint generation model as described above, with reference to... Figure 2 and Figure 3 As shown, it includes the following steps: S1. Obtain the metasurface array and form different encoding array structures according to the encoding state of the metasurface units, corresponding to different RCS responses; S2. Simulation datasets are generated by obtaining simulation RCS data of each coding array through electromagnetic simulation software CST. Simultaneously, calculation datasets are generated by combining theoretical RCS data calculated based on remote superposition formula. Formula-simulation residual network is constructed using calculation datasets and simulation datasets, and calculation datasets and simulation datasets are expanded respectively. S3. Establish a physical constraint depth condition generation model, which includes three parts: a prediction model, an identification model, and a generation model. S4. Use the expanded computational dataset to pre-train the prediction model, establish and quickly fit the physical constraint relationship to form a pre-trained prediction model. Then, use the expanded simulation dataset to jointly train the pre-trained prediction model, recognition model and generation model to obtain the trained model, and realize the parameter co-optimization and response consistency physical constraints between models. S5. Input the conditions of the target RCS into the trained model, output the optimal metasurface unit arrangement scheme, and the mechanical control device controls the metasurface units according to the optimal metasurface unit arrangement scheme to realize the rapid simulation and reconstruction of RCS.

[0023] Furthermore, the simulation dataset was used to simulate the full-space scattering characteristics of different metasurface coded arrays using the electromagnetic simulation software CST, obtaining radar cross section (RCS) data within a predetermined frequency band and angle range. The dataset is computed based on the pattern function of the metasurface to calculate the far-field scattering response with arbitrary phase distribution. The far-field superposition formula is used to theoretically calculate the corresponding theoretical RCS distribution for each coded array. The pattern function of the coded metasurface is used as the basis for this calculation. as follows: , The metasurface is an N×N array, and the size of each unit is... , The abrupt phase change provided by the cell located in the m-th row and n-th column for the j-th operating frequency point is: and These are pitch and azimuth angles. It is the wavenumber at the operating frequency.

[0024] Furthermore, refer to Figure 4 As shown, a formula-simulation residual network is constructed to augment the dataset using computational and simulation datasets. The residual network structure is a one-dimensional convolutional neural network, used to extrapolate simulation-like calibration data from the formula data at a lower cost, and to pre-train the prediction model. The prediction model, based on the convolutional neural network structure, is used to realize the positive mapping relationship between the metasurface array structure and its corresponding radar cross section (RCS) distribution. The input data is a 1×N×N single-channel coded array diagram, representing the array coded arrangement composed of N×N metasurface units, where each unit corresponds to either "0" or "1". The network consists of three sets of convolutional and pooling layers and two fully connected layers. The output data is a 181-dimensional RCS prediction result, corresponding to RCS values ​​from incident angle θ of -90° to +90°. Pre-training is performed using the computational dataset, and the root mean square error (RMSE) between the output RCS prediction value and the true RCS value is used as the loss function for optimization. The loss function adopts reconstruction loss. The expression is: , in, This is the true RCS value. To predict the RCS value, N is the number of sample points.

[0025] Furthermore, the prediction model is based on a convolutional neural network structure, used to extract features from the input metasurface encoding and output the corresponding RCS distribution prediction results; the recognition model is based on an encoding network structure, taking the metasurface encoding and the target RCS distribution as input, outputting the mean and variance parameters of the latent variables, and generating latent variable representations based on reparameterized sampling; the generation model is based on a decoding network structure, taking the target RCS distribution and latent variables as input, outputting the generated metasurface encoding scheme, realizing reverse design.

[0026] Furthermore, refer to Figure 5 As shown, considering the metasurface coupling effect, to enhance the stability of its unit structure, a 4×4 metasurface unit is regarded as a metasurface sub-unit group and set to the same encoding state. Therefore, the input of the recognition model includes the metasurface encoding vector N×N / 16 and the 181-dimensional RCS distribution vector, which are concatenated to form a 197-dimensional input vector. The network consists of two fully connected layers and two output layers. The first fully connected layer maps the input dimension from 197 to 512 neurons and applies ReLU activation. The second fully connected layer further reduces the dimension of the features to 256 neurons. Then, the latent variable z is output. The latent variable is used to characterize the implicit statistical relationship between the metasurface structure and the RCS response, providing randomness and controllability conditions for the subsequent generation model. The input to the generative model includes a 181-dimensional target RCS and a 64-dimensional latent variable vector z, which are concatenated to form a 245-dimensional input feature. The network consists of two fully connected layers and one output layer. The first layer maps the input dimension from 245 to 256 neurons and applies ReLU activation. The second layer maintains the dimension of 256 and continues to use ReLU activation to enhance nonlinear expressive power. The final output layer generates metasurface encoding and automatically outputs multiple sets of candidate metasurface array arrangement schemes under the constraint of the target RCS, realizing the inverse mapping process from electromagnetic response to structural design.

[0027] Furthermore, the total loss value of joint training is composed of the reconstruction loss. KL divergence regularization Coding reconstruction error With cycle consistency The composition, expressed as: , Among them, reconstruction loss It is the error between the RCS predicted by the prediction model and the true RCS, KL divergence regularization It is the KL divergence between the latent variable distribution and the standard normal distribution, and the encoding reconstruction error. It is the binary cross-entropy between the generated model's output encoding and the actual encoding, and cycle consistency. It is the consistency between the RCS predicted by the generative model and the true RCS after the generative model generates the encoded input and the predicted model. In other words, it is the function obtained by taking the difference between the predicted target RCS and the RCS predicted by the prediction model. This is then input into the model training process to give the model physical constraints. , , and These are the weighting coefficients for the four losses.

[0028] Furthermore, the optimal metasurface unit arrangement scheme includes: The generative model samples the target RCS distribution and outputs multiple candidate metasurface coding schemes. The candidate codes are input into the prediction model to obtain the corresponding RCS prediction results; the mixed similarity index between each prediction result and the target RCS is calculated. The mixed similarity includes the root mean square error (RMSE) and the structural similarity of the center angle range. The candidate scheme with the highest hybrid similarity is selected as the optimal metasurface encoding arrangement scheme.

[0029] Furthermore, in the mechanical control device, the control module is used to generate and output digital control signals according to the optimal metasurface coding arrangement scheme; The drive circuit is used to receive digital control signals and convert them into analog drive signals, and amplify the current to drive the stepper motor; A stepper motor is used to respond to the current signal of the drive circuit to drive the metasurface unit to rotate, so that the metal pattern layer on the top of the metasurface unit rotates to different orientation angles, thereby realizing the switching of the metasurface unit state.

[0030] In one specific embodiment, refer to Figure 6 As shown, , , and Using values ​​of 1.0, 0.1, 1.0, and 0.5 respectively, comparison charts were generated of the target RCS, model-predicted RCS, and model output encoding simulation results corresponding to the physical constraint depth conditions mentioned above. The simulation results corresponding to the model-predicted RCS and model output encoding were highly similar, proving the superior predictive ability of the prediction model. The Pearson correlation coefficient between the target RCS and the model-predicted RCS was 87.68%, meeting the simulation requirements.

[0031] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A metasurface control device based on a physical constraint generation model, characterized in that, Including metasurface arrays and mechanical control devices, The metasurface array consists of N×N metasurface units, each of which includes a bottom square metal layer, a middle cylindrical dielectric layer, and a top metal pattern layer. The mechanical control device includes a control module, a drive circuit, and stepper motors, with each stepper motor connected to a metasurface unit.

2. The metasurface control device based on a physical constraint generation model according to claim 1, characterized in that, The top metal pattern layer includes two circular I-shaped structures with the same opening angle but different orientation angles, which enable the metasurface units to form two encoding states of "0" and "1". Stepper motors are used to drive metasurface units to switch between "0" and "1" states.

3. A metasurface control method based on a physical constraint generation model, applied to the metasurface control device based on a physical constraint generation model as described in any one of claims 1-2, characterized in that, Includes the following steps: S1. Obtain the metasurface array and form different encoding array structures according to the encoding state of the metasurface units, corresponding to different RCS responses; S2. Simulation datasets are generated by obtaining simulation RCS data of each coding array through electromagnetic simulation software CST. Simultaneously, calculation datasets are generated by combining theoretical RCS data calculated based on remote superposition formula. Formula-simulation residual network is constructed using calculation datasets and simulation datasets, and calculation datasets and simulation datasets are expanded respectively. S3. Establish a physical constraint depth condition generation model, which includes three parts: a prediction model, an identification model, and a generation model. S4. Use the expanded computational dataset to pre-train the prediction model, establish and quickly fit the physical constraint relationship to form a pre-trained prediction model. Then, use the expanded simulation dataset to jointly train the pre-trained prediction model, recognition model and generation model to obtain the trained model, and realize the parameter co-optimization and response consistency physical constraints between models. S5. Input the conditions of the target RCS into the trained model, output the optimal metasurface unit arrangement scheme, and the mechanical control device controls the metasurface units according to the optimal metasurface unit arrangement scheme to realize the rapid simulation and reconstruction of RCS.

4. The metasurface control method based on a physical constraint generation model according to claim 3, characterized in that, The simulation dataset uses the electromagnetic simulation software CST to simulate the full-space scattering characteristics of different metasurface coded arrays, obtaining radar cross section (RCS) data within a predetermined frequency band and angle range; The dataset is computed based on the pattern function of the metasurface to calculate the far-field scattering response with arbitrary phase distribution. The far-field superposition formula is used to theoretically calculate the corresponding theoretical RCS distribution for each coded array. The pattern function of the coded metasurface is used as the basis for this calculation. as follows: ， The metasurface is an N×N array, and the size of each unit is... , The abrupt phase change provided by the cell located in the m-th row and n-th column for the j-th operating frequency point is: and These are pitch and azimuth angles. It is the wavenumber at the operating frequency.

5. The metasurface control method based on a physical constraint generation model according to claim 3, characterized in that, The prediction model is based on a convolutional neural network (CNN). It utilizes computational and simulation datasets to construct a formula-simulation residual network to augment the dataset. The residual network is a one-dimensional CNN used to extrapolate formula data to obtain simulation-like calibration data, which is then used for pre-training the prediction model. The root mean square error (RMSE) between the predicted and actual RCS values ​​is used as the loss function for optimization. The loss function employs a reconstruction loss mechanism. The expression is: ， in, This is the true RCS value. To predict the RCS value, N is the number of sample points.

6. The metasurface manipulation method based on a physical constraint generation model according to claim 3, characterized in that, The total loss value of joint training consists of reconstruction loss. KL divergence regularization Coding reconstruction error With cycle consistency The composition, expressed as: ， Among them, reconstruction loss It is the error between the RCS predicted by the prediction model and the true RCS, KL divergence regularization It is the KL divergence between the latent variable distribution and the standard normal distribution, and the encoding reconstruction error. It is the binary cross-entropy between the generated model's output encoding and the actual encoding, and cycle consistency. The consistency between the RCS predicted by the generative model and the true RCS after the generative model generates coded input and predicts the model gives the model physical constraints. , , and These are the weighting coefficients for the four losses.

7. The metasurface control method based on a physical constraint generation model according to claim 3, characterized in that, The optimal metasurface element arrangement scheme includes: The generative model samples the target RCS distribution and outputs multiple candidate metasurface coding schemes. The candidate codes are input into the prediction model to obtain the corresponding RCS prediction results; the mixed similarity index between each prediction result and the target RCS is calculated. The mixed similarity includes the root mean square error (RMSE) and the structural similarity of the center angle range. The candidate scheme with the highest hybrid similarity is selected as the optimal metasurface encoding arrangement scheme.

Images