A method and system for optimizing metasurface design based on task-oriented learning

Abstract

The present invention discloses a method and system for optimizing the design of a metasurface based on task-oriented learning, which relates to the technical field of electromagnetic metamaterial design and optimization. The method includes: generating the structural parameters of the metasurface unit and the corresponding electromagnetic characteristic parameters, and constructing a data set; constructing a neural network with the structural parameters of the unit as the input and the corresponding electromagnetic characteristic parameters as the output, and using the data set to train the neural network; defining the target characteristic parameters of the metasurface unit, and constructing an optimization objective function with the output parameters of the neural network and the target characteristic parameters as independent variables; taking the derivative of the objective function to obtain the adjustment amount of the structural parameters of the unit; updating the structural parameters of the unit; until the objective function meets the threshold, realizing the design optimization of the metasurface. The present invention provides a new idea for the development of metasurface technology through an intelligent optimization design method, and can meet the requirements for the multifunctional and efficient design of complex metasurfaces.

Description

Technical Field

[0001] The present invention relates to the technical field of electromagnetic metamaterial design and optimization, and particularly to a method and system for optimizing metasurface design based on task-oriented learning. Background Art

[0002] As a new type of artificial electromagnetic material, electromagnetic metasurface technology can precisely control the propagation behavior of electromagnetic waves by designing unit structures with sub-wavelength scales, including characteristics such as absorption, reflection, refraction, transmission, and polarization conversion. The emergence of metasurfaces provides a revolutionary solution for traditional electromagnetic wave control technologies. For example, by regulating the scattering characteristics of electromagnetic waves, metasurfaces can significantly reduce the radar cross-section (RCS) of target objects, thereby achieving stealth effects. Compared with traditional stealth materials, metasurface design has higher flexibility and can be customized according to the size, shape, and material of different targets, and is widely used in military protection, stealth aircraft, and unmanned aerial vehicle fields. In the field of wireless signal enhancement and beam control, metasurface technology can precisely control the propagation direction, intensity, and polarization of signals, thereby improving the signal coverage and reliability of communication systems. It shows unique advantages in beamforming and can achieve beam concentration or dispersion in a specified direction through the design of phase distribution. Compared with traditional phased array antennas, metasurface beamforming devices have the characteristics of simple structure, low cost, and light weight. In a complex electromagnetic environment, metasurface technology can be used as an absorbing material or shielding material to reduce electromagnetic wave interference and improve electromagnetic compatibility. Its high-efficiency wide-band absorption characteristics play an important role in scenarios such as radio frequency shielding, electronic device protection, and anti-electromagnetic attack.

[0003] However, the design of metasurfaces usually relies on experience and simulation optimization. Designers need to repeatedly adjust unit structure parameters (such as geometric dimensions, material parameters) to achieve the target performance. This process is time-consuming and laborious, especially when dealing with multi-variable optimization problems, the efficiency drops significantly. Existing design methods are difficult to quickly generate customized metasurface designs that meet different application scenarios in the face of complex scenarios, which limits the popularization of practical applications.

[0004] With the rapid development of machine learning technology, deep neural networks have demonstrated powerful capabilities in non-linear mapping modeling and optimization problems. By introducing deep learning methods into metasurface design and using neural networks to construct the non-linear mapping relationship between unit structure parameters and electromagnetic performance, the design efficiency can be significantly improved, and global optimization of complex electromagnetic performance can be achieved. However, the current metasurface design methods based on deep learning are still in the initial exploration stage, and there is still much room for improvement in multi-objective optimization, searching in complex parameter spaces, and adapting to different task requirements.

[0005] Therefore, it is an urgent problem for those skilled in the art to propose a method and system for optimizing metasurface design based on task-oriented learning to solve the difficulties existing in the prior art. Summary of the Invention

[0006] In view of this, the present invention provides a method and system for optimizing metasurface design based on task-oriented learning, which constructs the relationship between target performance and design parameters through a deep neural network, and combines the gradient descent method to quickly optimize the unit parameters, providing an efficient and intelligent solution for complex metasurface design.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A method for optimizing metasurface design based on task-oriented learning includes the following steps:

[0009] S1. Data generation: Generate the structural parameters p of the metasurface unit and the corresponding electromagnetic characteristic parameters s, and construct a data set And construct an empty corrected data set For later updating of network parameters, where p = {p1, p2,..., p i}, s = {s1, s2,..., s i}, and i represents the index of the sample;

[0010] S2. Network training: Construct a neural network φ θ with the structural parameters p of the unit as the input and the corresponding electromagnetic characteristic parameters s as the output, and use the data set

[0011] to train the neural network, where θ is the network parameter;

[0012] S3. Objective function construction: Define the target characteristic parameter s * of the metasurface unit, and construct an optimization objective function f(s * * , s) with the output parameter s of the neural network and the target characteristic parameter s as independent variables;

[0013] S4. Obtain the adjustment amount: Take the derivative of f(s * , s) to obtain the adjustment amount Δp of the structural parameters p of the unit,

[0014] S5. Update parameters: Use the adjustment amount Δp to update the structural parameters of the unit;

[0015] S6. Design optimization: Repeat S3 - S5 until the objective function meets the threshold to achieve the design optimization of the metasurface.Optionally, the metasurface unit structure parameters in S1 include side length, sheet resistance value, material sheet resistance, unit period length, dielectric constant of the material, and dielectric thickness.

[0016] Optionally, the neural network in S2 includes an input layer, a fully connected layer, a reshaping layer, a convolutional layer, a max pooling layer, and an output layer connected in sequence.

[0017] Optionally, the optimization objective function f(s * , s) has the following calculation formula:

[0018]

[0019] Optionally, in S4, the gradient descent method is used to iteratively adjust the unit structure parameters. The derivative of f(s * , s) is taken to obtain the adjustment amount Δp of the unit structure parameter p:

[0020]

[0021] where θ represents taking the derivative, and φ θ represents the parameter characteristic prediction neural network with parameter θ;

[0022] The unit structure parameters are updated according to the following formula:

[0023] p = p + ∈Δp

[0024] where 0 < ∈ ≤ 1 represents the update step size.

[0025] Optionally, in S6, S3 to S5 are repeated until the threshold satisfied by the objective function is as follows:

[0026] f(s * , s) < f T

[0027] where f T is the threshold of the optimization objective function.

[0028] Optionally, if f(s * , s) < f T holds, then an electromagnetic simulation software is used for simulation to obtain the electromagnetic characteristic parameters corresponding to the structure parameter p at this moment. Then the objective function at this moment is Calculate the objective function at this moment. And put the training sample into the correction data set.

[0029] If then stop the iteration; otherwise, return to S4 to continue the update iteration.

[0030] Optionally, the optimizers adopted during the neural network training process include Adam and RMSprop.

[0031] A metasurface design optimization system based on task-oriented learning, applying the metasurface design optimization method based on task-oriented learning according to any one of the above, includes: a data generation module, a network training module, an objective function construction module, an adjustment amount acquisition module, a parameter update module, and a design optimization module;

[0032] The data generation module, connected to the input end of the network training module, is used to generate the structural parameters of the metasurface unit and the corresponding electromagnetic characteristic parameters, and construct a data set and a corrected data set;

[0033] The network training module, connected to the input end of the objective function construction module, is used to construct a neural network with the unit structural parameters as the input and the corresponding electromagnetic characteristic parameters as the output, and use the data set to train the neural network;

[0034] The objective function construction module, connected to the input end of the adjustment amount acquisition module, is used to define the target characteristic parameters of the metasurface unit and construct an optimization objective function with the output parameters of the neural network and the target characteristic parameters as independent variables;

[0035] The adjustment amount acquisition module, connected to the input end of the parameter update module, is used to take the derivative of the objective function to obtain the adjustment amount of the unit structural parameters;

[0036] The parameter update module, connected to the input end of the design optimization module, is used to update the unit structural parameters by using the adjustment amount;

[0037] The design optimization module is used to repeat the construction of the objective function, the acquisition of the adjustment amount, and the update of the parameters until the objective function meets the threshold, so as to realize the design optimization of the metasurface.

[0038] It can be seen from the above technical solutions that, compared with the prior art, the present invention provides a metasurface design optimization method and system based on task-oriented learning, and has the following beneficial effects: The present invention provides a new idea for the development of metasurface technology through an intelligent optimization design method, and can meet the requirements for the multi-functional and efficient design of complex metasurfaces. Description of the Drawings

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained according to the provided drawings without creative efforts.

[0040] Figure 1Flowchart of a method for optimizing metasurface design based on task - oriented learning provided by the present invention;

[0041] Figure 2 Test result graph of the indoor experimental flexible stealth device provided by the embodiment of the present invention, where 2a is the wave absorption rate in simulation and experiment, and 2b shows the wave absorption rate at different incident wave angles. Detailed implementation manners

[0042] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0043] Refer to Figure 1 As shown, the present invention discloses a method for optimizing metasurface design based on task - oriented learning, which can be widely applied to technical directions such as radar stealth, beam control, wireless signal enhancement, and electromagnetic environment management, and is particularly suitable for optimizing the structural parameters of complex multi - variable metasurface units, aiming to achieve fast and efficient metasurface design and performance optimization, including the following steps:

[0044] S1. Data generation: Generate the structural parameters p of the metasurface unit and the corresponding electromagnetic characteristic parameters s, and construct a data set And construct an empty corrected data set For later updating of network parameters, where p = {p1, p2,..., p i}, s = {s1, s2,..., s i}, and i represents the index of the sample;

[0045] S2. Network training: Construct a neural network φ θ with the structural parameters p of the unit as the input and the corresponding electromagnetic characteristic parameters s as the output, and use the data set to train the neural network, where θ is the network parameter;

[0046] S3. Objective function construction: Define the target characteristic parameters s * of the metasurface unit, and construct an optimization objective function f(s * , s) with the output parameter s of the neural network and the target characteristic parameter s * as independent variables;

[0047] S4. Obtain the adjustment amount: Take the derivative of f(s * , s) to obtain the adjustment amount Δp of the structural parameters p of the unit,

[0048] S5. Update parameters: Update the unit structure parameters using the adjustment amount Δp;

[0049] S6. Design optimization: Repeat S3 - S5 until the objective function meets the threshold to achieve the design optimization of the metasurface.

[0050] Furthermore, the metasurface unit structure parameters in S1 include side length, sheet resistance value, material sheet resistance, unit period length, dielectric constant of the material, and dielectric thickness.

[0051] Specifically, by optimizing these structure parameters, ensure that the metasurface unit achieves high absorption efficiency within the target frequency range (such as 9 - 25 GHz), while having angular insensitivity and polarization - independent characteristics.

[0052] Furthermore, the neural network in S2 includes an input layer, a fully - connected layer, a reshaping layer, a convolutional layer, a max - pooling layer, and an output layer connected in sequence.

[0053] Furthermore, the calculation formula for optimizing the objective function f(s * , s) in S3 is:

[0054]

[0055] Furthermore, in S4, the gradient - descent method is used to iteratively adjust the unit structure parameters, and the derivative of f(s * , s) is taken to obtain the adjustment amount Δp of the unit structure parameter p:

[0056]

[0057] Among them, denotes taking the derivative, and φ θ denotes the parameter - characteristic prediction neural network with parameter θ;

[0058] The formula for updating the unit structure parameters is as follows:

[0059] p = p + ∈Δp

[0060] Among them, 0 < ∈ ≤ 1 represents the update step size.

[0061] Furthermore, the threshold for repeating S3 - S5 until the objective function is satisfied in S6 is as follows:

[0062] f(s * , s) < f T

[0063] Among them, f T is the threshold of the optimization objective function.

[0064] Furthermore, if f(s * , s) < f TIf it holds, use electromagnetic simulation software for simulation to obtain the electromagnetic characteristic parameters corresponding to the structural parameter p at this moment. Then the objective function at this moment is Calculate the objective function at this moment And put the training samples Into the correction data set

[0065] If Then stop the iteration; otherwise, return to S4 to continue the update and iteration.

[0066] Furthermore, the optimizers adopted during the neural network training process include Adam and RMSprop.

[0067] In a specific embodiment, a meta - surface design optimization method based on task - oriented learning is applied to a stealth device. The stealth device is designed based on a flexible substrate and a deformable meta - surface unit structure, which can adapt to the surface of complex and irregular target objects to achieve a flexible stealth effect. The stealth device can be shaped into various shapes, such as an invisible tent, to shield internal objects from external radar (such as UAV SAR radar) detection.

[0068] Apply a flexible leather substrate and a stretchable conductive ink layer to make the stealth device have excellent ductility. By optimizing the key parameters, the target electromagnetic performance (such as broadband high absorption efficiency) can be achieved.

[0069] The key parameters are specifically: conductive pattern: side lengths l1, l2, sheet resistance values m, n; substrate parameters: dielectric thicknesses h1, h2.

[0070] Construct a neural network, with the input being the unit structure parameters [h1, h2, l1, l2, m, n] and the output being the S - parameters. Construct an adaptive objective function f(s * , s), which represents the error between the s - parameter s of the current structure and the target s - parameter s * Then, based on the gradient descent method, optimize the objective function and adjust the structural parameters until the condition is satisfied, where f T is the convergence threshold.

[0071] The pseudo - code of the task - oriented learning (TOL) algorithm is as follows:

[0072]

[0073]

[0074] To fit the function, an embodiment of the present invention proposes a neural network consisting of six layers. The input of the neural network includes six variables. The middle three layers consist of fully connected layers with 256 units and 2048 units. Then, the 2048 linear outputs are reshaped into a two-channel 32x32 image. Next, 16 convolutional filters of size 5x5 with a stride of 2 are applied to extract features, followed by a 2x2 pooling layer. The subsequent layer is a convolutional layer with a 3x3 filter and a stride of 1. Finally, two fully connected layers with 512 units and 256 units respectively are applied, and 256 S-parameter values are output. The network structure parameters are summarized in Table 1.

[0075] Table 1 Network Structure Parameters

[0076] Layer type Detailed description Input layer <![CDATA[Input: [h1, h2, l1, l2, m, n] (6 variables)]]> Fully connected layer 256 units, fully connected Fully connected layer 512 units, fully connected Reshape layer Reshape 256 units into a 16x16 image with 2 channels Convolutional layer 16 filters, 5x5 convolutional kernel, stride 2 Max pooling layer 3x3 pooling Convolutional layer 8 filters, 3x3 convolutional kernel, stride 1 Fully connected layer 512 units, fully connected Output layer 256 units, fully connected

[0077] When training the network parameters, simulated data generated using the Python and CST interfaces is used. Finally, a dataset containing 20,000 continuous samples is generated. During the training process, the mean squared error (MSE) is used as the loss function, defined as where N B represents the number of samples in each batch, represents the predicted value, s i represents the actual value. To efficiently train the network, the classical ADAM optimizer is used. The training and validation losses converge rapidly at the 400th epoch. Figure 2 The test results of the indoor experimental flexible stealth device of this embodiment are shown, where 2a is the absorption rate in simulation and experiment, and 2b shows the absorption rate at different incident wave angles.

[0078] After the stealth device is prepared, its absorption performance is tested in a microwave anechoic chamber using a separated radar transmitting and receiving system. The results show that the absorption rate exceeds 90% in the frequency range of 9–25 GHz and exhibits good angle-insensitive characteristics in the incident angle range of 0°–60°.

[0079] Specifically, the structural similarity index (SSIM) is a method for evaluating the difference in structural information between images. Different from traditional pixel-by-pixel error measurement methods, SSIM evaluates images based on three aspects: brightness, contrast, and structure, providing a more comprehensive image quality assessment.

[0080] The brightness, contrast, and structure-related factors are defined as l(x,y), c(x,y), and s(x,y) respectively, and are considered to be independent of each other. The SSIM index can be defined using the following functions S1 and S2:

[0081] SSIM(x,y)) = f(l(x,y), c(x,y), s(x,y))

[0082] The derivation of the brightness - related factor is as follows:

[0083]

[0084] Among them, μ x and μ y are the average values of the SAR image data before and after protection respectively. C1 is a small constant to avoid division by zero, which is set small enough to satisfy the Weber - Fechner law, usually C1=(K1L) 2 .

[0085] The calculation of the contrast - related factor uses the following function:

[0086]

[0087] Among them, σ x 2 and σ y 2 represent the variances of the SAR image data before and after protection respectively. C2 is defined as C2=(K2L) 2 .

[0088] Structure - related factor:

[0089]

[0090] Among them, σ xy is the covariance of the SAR image data before and after protection, and C3 = C2 / 2.

[0091] In summary, the structural similarity index integrates these factors to provide a measure of the structural similarity between images as follows:

[0092] SSIM(x,y))=[l(x,y)] α ·[c(x,y)] β ·[s(x,y)] γ

[0093] Among them, α, γ and β are adjustable parameters that determine the importance of each similarity factor. The range of the SSIM index is [- 1,1], where a value of 1 means that the two images are exactly the same.

[0094] Experimental results: When the stealth device is not covered, the scattering intensity of the target object in the SAR image is obvious and easy to be recognized by the radar. After covering the stealth device, the scattering intensity of the target object is significantly reduced and almost merged with the background. Extracting the target - area image for SSIM analysis, the results show that the stealth device effectively suppresses the radar signal of the target object.

[0095] Through the above embodiments, the present invention realizes the metasurface design of a SAR stealth device with a wide bandwidth and high absorption efficiency, providing a feasible solution for the practical application of the metasurface design optimization method based on task-oriented learning.

[0096] Corresponding to Figure 1 the method described above, an embodiment of the present invention further provides a metasurface design optimization system based on task-oriented learning, specifically including: a data generation module, a network training module, a target function construction module, an adjustment amount acquisition module, a parameter update module, and a design optimization module;

[0097] The data generation module is connected to the input end of the network training module, and is used to generate the structural parameters of the metasurface unit and the corresponding electromagnetic characteristic parameters, and construct a data set and a corrected data set;

[0098] The network training module is connected to the input end of the target function construction module, and is used to construct a neural network with the unit structural parameters as the input and the corresponding electromagnetic characteristic parameters as the output, and use the data set to train the neural network;

[0099] The target function construction module is connected to the input end of the adjustment amount acquisition module, and is used to define the target characteristic parameters of the metasurface unit, and construct an optimization target function with the output parameters of the neural network and the target characteristic parameters as independent variables;

[0100] The adjustment amount acquisition module is connected to the input end of the parameter update module, and is used to take the derivative of the target function to obtain the adjustment amount of the unit structural parameters;

[0101] The parameter update module is connected to the input end of the design optimization module, and is used to update the unit structural parameters by using the adjustment amount;

[0102] The design optimization module is used to repeat the construction of the target function, the acquisition of the adjustment amount, and the update of the parameters until the target function meets the threshold, so as to realize the design optimization of the metasurface.

[0103] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The same or similar parts among the various embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the description of the method part.

[0104] The foregoing description of the disclosed embodiments enables those skilled in the art to practice or use the present 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 present invention. Thus, the present invention is not intended 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 method for optimizing the design of a metasurface based on task-oriented learning, characterized in that, It includes the following steps: S1. Data generation: Generate the structural parameters p of the metasurface unit and the corresponding electromagnetic characteristic parameters s, and construct a data set And construct an empty corrected data set For later updating of network parameters, where p = {p1, p2,..., p i}, s = {s1, s2,..., s i}, and i represents the index of the sample; S2. Network training: Construct a neural network φ with the unit structure parameter p as the input and the corresponding electromagnetic characteristic parameter s as the output θ , and use the dataset to train the neural network, where θ is the network parameter; S3. Objective function construction: Define the target feature parameter s of the metasurface unit * , construct an optimization objective function f(s * , s) with the output parameter s of the neural network and the target feature parameter s * as independent variables; S4. Obtain the adjustment amount: Differentiate f(s * , s) to obtain the adjustment amount Δp of the unit structure parameter p. S5. Update parameters: Update the unit structure parameters by using the adjustment amount Δp; S6. Design optimization: Repeat S3 - S5 until the objective function meets the threshold to achieve the design optimization of the metasurface; The calculation formula of the optimized objective function f(s * , s) in S3 is as follows: In S4, the gradient descent method is used to iteratively adjust the unit structure parameters, and the derivative of f(s * , s) is calculated to obtain the adjustment amount Δp of the unit structure parameter p: Among them, represents taking the derivative, and φ θ represents the parametric characteristic prediction neural network with parameter θ; The update of the unit structure parameters follows the following formula: p = p + ∈Δp where 0 < ∈ ≤ 1 represents the update step size.

2. A metasurface design optimization method based on task - oriented learning according to claim 1, characterized in that the unit structure parameters of the metasurface in S1 include side length, sheet resistance value, material sheet resistance, unit period length, dielectric constant of the material, and dielectric thickness.

3. A metasurface design optimization method based on task - oriented learning according to claim 1, characterized in that the neural network in S2 includes an input layer, a fully - connected layer, a reshaping layer, a convolutional layer, a max - pooling layer, and an output layer connected in sequence.

4. A metasurface design optimization method based on task - oriented learning according to claim 1, characterized in that the threshold that S6 repeats S3 - S5 until the objective function meets is as follows: f(s * ,s)<f T Among them, f T is the threshold of the optimization objective function.

5. A metasurface design optimization method based on task - oriented learning according to claim 1 or 4, characterized in that If f(s * , s) < f T holds, then an electromagnetic simulation software is used for simulation to obtain the electromagnetic characteristic parameters corresponding to the structural parameter p at this moment Then the objective function at this moment is Calculate the objective function at this moment And put the training sample into the correction data set If Stop the iteration; otherwise, return to S4 to continue updating and iterating.

6. A metasurface design optimization method based on task - oriented learning according to claim 1, characterized in that the optimizers used in the neural network training process include Adam and RMSprop.

7. A meta-surface design optimization system based on task-oriented learning, characterized in that Applying the metasurface design optimization method based on task - oriented learning according to any one of claims 1 - 6 includes: a data generation module, a network training module, an objective function construction module, an adjustment amount acquisition module, a parameter update module, and a design optimization module; The data generation module, connected to the input end of the network training module, is used to generate the unit structure parameters of the metasurface and the corresponding electromagnetic characteristic parameters, and construct a data set and a corrected data set; The network training module, connected to the input end of the objective function construction module, is used to construct a neural network with the unit structure parameters as the input and the corresponding electromagnetic characteristic parameters as the output, and use the data set to train the neural network; The objective function construction module, connected to the input end of the adjustment amount acquisition module, is used to define the target characteristic parameters of the metasurface unit and construct an optimization objective function with the output parameters of the neural network and the target characteristic parameters as independent variables; The adjustment amount acquisition module, connected to the input end of the parameter update module, is used to take the derivative of the objective function to obtain the adjustment amount of the unit structure parameters; The parameter update module, connected to the input end of the design optimization module, is used to update the unit structure parameters by using the adjustment amount; The design optimization module is used to repeat the objective function construction, adjustment amount acquisition, and parameter update until the objective function meets the threshold to achieve the design optimization of the metasurface; The calculation formula for optimizing the objective function f(s * , s) is as follows: Use the gradient descent method to iteratively adjust the unit structure parameters, and take the derivative of f(s * , s) to obtain the adjustment amount Δp of the unit structure parameter p: Among them, denotes taking the derivative, and φ θ denotes the parametric characteristic prediction neural network with parameter θ; The update of the unit structure parameters follows the following formula: p = p + ∈Δp where 0 < ∈ ≤ 1 represents the update step size.

Images