Deep learning-based broadband metamaterial wave-absorbing structure design method

Abstract

The invention discloses a broadband metamaterial wave-absorbing structure design method based on deep learning, a metal-medium-metal metamaterial wave-absorbing structure is constructed, the bottom layer is a metal layer, the middle layer is a medium layer, the top layer is a metal resonance layer, the metal resonance layer is composed of two parts, namely a square ring with two identical notches, the notches are symmetrical about a diagonal line, and the metal resonance layer is a square ring. The central area is composed of small metal square blocks, and the central area composed of the small square blocks is surrounded by a square ring with a notch; a binary (0, 1) matrix is randomly generated to represent the positions of the small metal blocks in the central area, and a training data set containing the binary (0, 1) matrix and electromagnetic response parameters is constructed; constructing and training a forward prediction convolutional neural network by taking the binary (0, 1) matrix as input and the electromagnetic response parameters as output; taking the metamaterial electromagnetic response parameter and the bandwidth ratio as input, taking a binary (0, 1) matrix as output, and constructing and training a reverse design deep convolutional neural network; according to the invention, the metamaterial wave-absorbing structure can be rapidly designed.

Description

Technical Field

[0001] This application relates to the design of microwave absorption structures, and particularly to a method for designing broadband metamaterial microwave absorption structures based on deep learning. Background Art

[0002] Absorbing metamaterials have multi-faceted advantages over traditional microwave absorption materials, such as thin thickness, rich electric loss mechanisms, and lower density. Due to these advantages, absorbing metamaterials have become one of the research hotspots in recent years. Traditionally, people usually use a combination of finite element modeling and finite difference time domain and iterative calculation methods for the design of metamaterial structures. However, with the increase in structural complexity, it takes a long time to predict the properties of new metamaterial structures using this method, and it is easily affected by the experimental conditions and theoretical knowledge of the designer. The design space and design effect are limited, and it is difficult for pure manpower to complete a large amount of data analysis. Moreover, the metamaterial structures designed manually are easily trapped in the dilemma of local optimum. Summary of the Invention

[0003] The purpose of this application is to propose a method for designing broadband metamaterial microwave absorption structures based on deep learning, which can quickly design metamaterial microwave absorption structures by establishing the mapping relationship between the structural parameters of metamaterials and electromagnetic responses.

[0004] The technical solution to realize this application includes the following steps: A method for designing broadband metamaterial microwave absorption structures based on deep learning includes the following steps:

[0005] Step 1: Construct a metamaterial microwave absorption structure of metal-dielectric-metal, with a metal layer at the bottom, a dielectric layer in the middle, and a metal resonant layer at the top. The metal resonant layer consists of two parts. One part is a square ring with two identical notches, and the notches are symmetric about the diagonal. The other part is the central region composed of metal small squares, and the square ring with notches surrounds the central region composed of small squares.

[0006] Step 2: Randomly generate a binary (0,1) matrix to represent the positions of the metal small squares in the central region, where 1 represents the presence of a metal small square and 0 represents the absence. Conduct electromagnetic simulation to construct a training dataset containing the binary (0,1) matrix and electromagnetic response parameters.

[0007] Step 3: Use the binary (0,1) matrix as the input and the corresponding electromagnetic response parameters as the output to construct a forward prediction convolutional neural network to realize the spectrum prediction of the metamaterial microwave absorption structure.

[0008] Step 4: Use the electromagnetic response parameters and bandwidth ratio of the metamaterial as the input and the corresponding binary (0,1) matrix as the output to construct a reverse design deep convolutional neural network to realize the design from the expected electromagnetic response parameters to the metamaterial structure.

[0009] Step 5: Set the loss function and optimizer, and train the forward prediction network model and the inverse design deep convolutional neural network.

[0010] Further, in Step 2: Randomly generate a binary (0,1) matrix to represent the positions of the metal squares in the central region, where 1 represents the presence of a metal square and 0 represents the absence. Conduct electromagnetic simulation to construct a training dataset containing the binary (0,1) matrix and electromagnetic response parameters. The specific method is as follows:

[0011] Randomly generate a binary (0,1) matrix to represent the positions of the metal squares in the central region, where 1 represents the presence of a metal square and 0 represents the absence;

[0012] Conduct electromagnetic simulation, directly export the S11 parameters through CST. The S11 parameters usually have one thousand frequency response points. Uniformly select one point at intervals of 0.5 GHz as the electromagnetic response parameter.

[0013] Further, in Step 3: Construct a forward prediction convolutional neural network with the binary (0,1) matrix as the input and the corresponding electromagnetic response parameters as the output to achieve the spectrum prediction of the metamaterial absorbing structure, where:

[0014] In the forward prediction convolutional neural network, the first layer consists of 16 convolutional kernels of size 3×3, the second layer consists of 2 convolutional kernels of size 3×3, the third layer consists of a pooling layer of size 2×2, the fourth and fifth layers both consist of 2 convolutional layers of size 3×3, the sixth layer is a Flatten layer, the seventh and eighth layers are fully connected layers with the activation function Relu, and the last layer is a fully connected layer with the activation function Softmax.

[0015] Further, in Step 4: Construct an inverse design deep convolutional neural network with the metamaterial electromagnetic response parameters and bandwidth ratio as the input and the corresponding binary (0,1) matrix as the output to achieve the design from the desired electromagnetic response parameters to the metamaterial structure. The specific method is as follows:

[0016] The first, second, and third layers of the inverse design network are all fully connected layers with the activation function Relu. The fourth and fifth layers are convolutional layers composed of a convolutional kernel of size 3×3. The sixth layer is a pooling layer of size 2×2. The seventh and eighth layers are convolutional layers composed of 2 convolutional kernels of size 3×3. The ninth layer is a pooling layer of size 2×2. The tenth and eleventh layers are convolutional layers composed of convolutional kernels of size 3×3. The twelfth layer is a normalization BN layer. The thirteenth layer is a convolutional layer with a convolutional kernel size of 3×3. The fourteenth layer is a fully connected layer with the activation function Relu. The fifteenth layer is a fully connected layer with the activation function tanh.

[0017] Further, Step 5: Set the loss function and the optimizer, and train the forward prediction network and the inverse design deep convolutional neural network, where:

[0018] During the process of training the forward prediction network, the loss function is defined as the mean square error MSE, and the Adam optimizer is used for training. The loss function of the trained network stabilizes at about 10 -5 or so.

[0019] Further, Step 5: Set the loss function and the optimizer, and train the forward prediction network model and the inverse design deep convolutional neural network, where:

[0020] During the process of training the inverse design deep convolutional neural network, cascade the inverse design deep convolutional neural network with the forward prediction network. The forward prediction network is connected to the output end of the inverse network. By minimizing the error between the input of the inverse network and the output of the forward prediction network, the loss function adopts the mean square error MSE. Finally, the loss function stabilizes at about 10 -4 or so.

[0021] A broadband metamaterial absorbing structure is constructed based on the above-mentioned broadband metamaterial absorbing structure design method of deep learning.

[0022] A computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the above-mentioned broadband metamaterial absorbing structure design method of deep learning to realize the design of the broadband metamaterial absorbing structure.

[0023] A computer-readable storage medium stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned broadband metamaterial absorbing structure design method of deep learning to realize the design of the broadband metamaterial absorbing structure.

[0024] Compared with the prior art, the remarkable advantage of the present invention is that the middle binary (0,1) matrix can be flexibly changed according to different usage requirements. At the same time, with the help of the trained neural network, the mapping process from the binary matrix to the electromagnetic parameters and from the electromagnetic parameters to the binary matrix can be quickly realized. Description of the Drawings

[0025] Figure 1 is a schematic diagram of a metal-dielectric-metal metamaterial absorbing structure.

[0026] Figure 2 is a structural diagram of the forward prediction deep convolutional neural network.

[0027] Figure 3 is the specific architecture of the inverse design deep convolutional neural network.

[0028] Figure 4 This is the first example figure of the present invention. The structural schematic diagram is shown in the lower right corner of the figure, where black represents the presence of metal blocks. This figure shows the comparison of the electromagnetic parameters of the wave-absorbing structure obtained by the forward prediction network and the electromagnetic parameters obtained by CST simulation.

[0029] Figure 5 This is the second example figure of the present invention. Among them, (a) the input desired bandwidth ratio is 0.87, and the structure given by the inverse design network; (b) the Python-CST co-simulation structure; (c) the electromagnetic response parameters actually simulated by CST using the given structure, the bandwidth ratio is 0.87, which is consistent with the expectation; (d) the red curve is the target result, and the black is the electromagnetic response diagram of the structure given by the inverse design. The inverse design network can meet the desired design goal. Specific implementation manners

[0030] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0031] A broadband metamaterial wave-absorbing structure design method based on deep learning includes the following steps:

[0032] Step 1, construct a miniaturized broadband metamaterial wave-absorbing structure. The metamaterial wave-absorbing structure adopts a metal-dielectric-metal metamaterial wave-absorbing structure unit, such as Figure 1 shown.

[0033] For the metal-dielectric-metal metamaterial wave-absorbing structure, the bottom layer is a metal layer, the middle layer is a dielectric layer, and the top layer is a metal resonance layer. The periphery of the metal resonance layer is composed of a square ring with two notches.

[0034] Parameterize the internal region of the resonance layer of the metamaterial wave-absorbing structure into metal squares of the same size, where 0 represents no filling and 1 represents a metal square. After such encoding, a (0,1) matrix can be obtained.

[0035] Step 2, randomly generate a binary (0,1) matrix, perform electromagnetic simulation, and construct a training data set;

[0036] Use a random binary dot matrix to perform electromagnetic simulation, and directly export the S11 parameter through CST. The S11 parameter usually has one thousand frequency response points. Select one point at a 0.5 GHz interval evenly as the electromagnetic response parameter.

[0037] Step 3, construct a forward prediction convolutional neural network with the binary (0,1) matrix as the input and the corresponding electromagnetic response parameter as the output to realize the spectrum prediction of the metamaterial wave-absorbing structure;

[0038] In the forward prediction convolutional neural network, the first layer consists of 16 convolutional kernels of size 3×3, the second layer consists of 2 convolutional kernels of size 3×3, the third layer consists of a pooling layer of size 2×2, the fourth and fifth layers both consist of 2 convolutional layers of size 3×3, the sixth layer is a Flatten layer, the seventh and eighth layers are fully connected layers with the activation function Relu, and the last layer is a fully connected layer with the activation function Softmax. The specific network architecture is as Figure 2 shown.

[0039] The loss function is defined as the mean squared error MSE, and the Adam optimizer is used for training. The loss function of the trained network stabilizes at about 10 -5 or so.

[0040] Step 4: Using the electromagnetic response parameters and bandwidth ratio of the metamaterial as inputs and the corresponding binary (0,1) matrix as the output, construct a reverse design deep convolutional neural network to achieve the design from the electromagnetic response parameters to the metamaterial structure.

[0041] The reverse network needs to be cascaded with the trained forward prediction deep convolutional neural network. At this time, the weights of the forward prediction network are fixed and remain unchanged. The input structure parameters (binary (0,1) matrix) can give the electromagnetic response parameters through the forward prediction network. The network trains by minimizing the error between the input of the reverse network and the output of the forward prediction network. After training, the reverse network is separated from the cascaded structure and used as an independent network.

[0042] The first, second, and third layers of the reverse design network are all fully connected layers with the activation function Relu. The fourth and fifth layers are convolutional layers composed of a convolutional kernel of size 3×3. The sixth layer is a pooling layer of size 2×2. The seventh and eighth layers are convolutional layers composed of 2 convolutional kernels of size 3×3. The ninth layer is a pooling layer of size 2×2. The tenth and eleventh layers are convolutional layers composed of convolutional kernels of size 3×3. The twelfth layer is a normalization BN layer. The thirteenth layer is a convolutional layer with a convolutional kernel size of 3×3. The fourteenth layer is a fully connected layer with the activation function Relu. The fifteenth layer is a fully connected layer with the activation function tanh. Finally, it is cascaded with the forward prediction convolutional neural network. After training and iteration, the MSE stabilizes at the 10 -4 level. The specific architecture of the reverse design network is as Figure 3 .

[0043] Example 1

[0044] This example provides an example of using a forward prediction neural network to predict the electromagnetic response parameters of a metamaterial absorber structure. A unit structure of a metal-dielectric-metal metamaterial absorber with a 16×16 (0,1) matrix in the middle is built. The frequency range is selected from 7 GHz to 25 GHz. The bottom layer is a copper substrate, and the middle dielectric layer uses SiO 2 material. The surface metal resonant layer uses ITO material. The conductivity of the outer ring ITO material is 258 S / m, and the conductivity of the middle part of ITO is 3000 S / m. The geometric width of the periodic unit is p = 3.7 mm. The thicknesses of the top resonant layer and the bottom metal material are t = 0.06 mm. The thickness of the dielectric layer SiO 2 is d = 2.8 mm. The length of the outer ring is L = 3.35 mm. The side length of the central coding area is C = 1.6 mm. The width of the ring is G = 0.13 mm. The total thickness of the metamaterial unit structure is 2.92 mm. As shown in Figure 4 the upper left is the schematic diagram of the unit structure. By inputting this 16×16 (0,1) matrix into the forward prediction neural network, the electromagnetic response parameters of the absorber structure corresponding to this structure are output, and the results are compared with the results obtained by CST simulation. As shown in Figure 4 the lower right is the schematic diagram of the central coding structure, where black represents the existence of metal blocks. It can be concluded that the electromagnetic parameters predicted by the network are almost the same as those obtained by simulation.

[0045] Through this example, the forward prediction neural network has a good fitting effect in predicting the electromagnetic response parameters of the metamaterial absorber structure and can be used as a forward network cascaded with the inverse design network.

[0046] Example Two

[0047] This example uses the unit structure of the metal-dielectric-metal metamaterial absorber structure described in Example One, aiming to obtain a broadband metamaterial absorber structure unit with a bandwidth ratio above 0.8 and a peak absorption rate exceeding 99%. To verify the effectiveness of the inverse network, random tests are carried out. For the inverse design network that isolates the cascaded model, the desired electromagnetic response is input, and its corresponding 16×16 binary (0,1) matrix is output. The inverse design result is as shown in Figure 5 (a) in the figure. The structural parameters given by the inverse network are used to build a metamaterial absorber structure through Python-CST joint simulation. After verification by CST simulation, the actual bandwidth ratio is consistent with the expectation, as shown in Figure 5 (c) in the figure, and the design result meets the expected goal.

[0048] In summary, the present invention realizes a design scheme of a broadband metamaterial absorbing structure based on deep learning. A network is built using a deep learning model, and a forward prediction network of the metamaterial absorbing structure based on a convolutional neural network is designed to fit the results of the electromagnetic simulation software CST. The cascade model method is used to build a reverse design network on the basis of the forward prediction network. The trained reverse design network can quickly give the parameters of the metamaterial absorbing structure with specified electromagnetic response parameters.

[0049] The above embodiments only express the implementation modes of the present application, and the description thereof is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several modifications and improvements can be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A method for designing a broadband metamaterial absorbing structure based on deep learning, characterized in that: The following steps are involved: Step 1: Construct a metal-dielectric-metal metamaterial absorbing structure, with a metal layer at the bottom, a dielectric layer in the middle, and a metal resonance layer at the top. The metal resonance layer consists of two parts. One part is a square ring with two identical notches, and the notches are symmetrical about the diagonal line. The other part is a central area composed of small metal squares. The square ring with notches surrounds the central area composed of small squares. Step 2: Randomly generate a binary (0,1) matrix to represent the position of the small metal cube in the central area, where 1 represents the presence of a small metal cube and 0 represents the absence of a small metal cube. Perform electromagnetic simulation to construct a training data set containing the binary (0,1) matrix and electromagnetic response parameters. Step 3: Using the binary (0,1) matrix as input and the corresponding electromagnetic response parameters as output, a forward prediction convolutional neural network is constructed to achieve spectrum prediction of the metamaterial absorbing structure; Step 4: Using the electromagnetic response parameters and bandwidth ratio of the metamaterial as input and the corresponding binary (0,1) matrix as output, a reverse design deep convolutional neural network is constructed to achieve the design from the desired electromagnetic response parameters to the metamaterial structure; Step 5: Set the loss function and optimizer, train the forward prediction network model and reversely design the deep convolutional neural network.

2. The method for designing broadband metamaterial absorbing structures based on deep learning according to claim 1, characterized in that: Step 2: Randomly generate a binary (0,1) matrix to represent the position of the small metal cube in the central area, where 1 represents the presence of a small metal cube and 0 represents the absence of a small metal cube. Perform electromagnetic simulation to construct a training data set containing a binary (0,1) matrix and electromagnetic response parameters. The specific method is as follows: A binary (0,1) matrix is ​​randomly generated to represent the position of the metal cube in the central area, where 1 represents the presence of a metal cube and 0 represents the absence of a metal cube; Perform electromagnetic simulation and directly derive the S11 parameters through CST. The S11 parameters usually have one thousand frequency response points, and one point is evenly selected at 0.5 GHz intervals as the electromagnetic response parameters.

3. The method for designing broadband metamaterial absorbing structures based on deep learning according to claim 1, characterized in that: Step 3: Using the binary (0,1) matrix as input and the corresponding electromagnetic response parameters as output, a forward prediction convolutional neural network is constructed to predict the spectrum of the metamaterial absorbing structure, where: In the forward prediction convolutional neural network, the first layer consists of 16 convolution kernels of size 3×3, the second layer consists of 2 convolution kernels of size 3×3, the third layer consists of a pooling layer of size 2×2, the fourth and fifth layers are composed of 2 convolution layers of size 3×3, the sixth layer is a Flatten layer, the seventh and eighth layers are fully connected layers with Relu activation function, and the last layer is a fully connected layer with Softmax activation function.

4. The method for designing broadband metamaterial absorbing structures based on deep learning according to claim 1, characterized in that: Step 4: Using the electromagnetic response parameters and bandwidth ratio of the metamaterial as input and the corresponding binary (0,1) matrix as output, a reverse design deep convolutional neural network is constructed to achieve the design from the desired electromagnetic response parameters to the metamaterial structure. The specific method is as follows: The first, second and third layers of the reverse-designed network are all fully connected layers with Relu as the activation function, the fourth and fifth layers are convolutional layers consisting of a 3×3 convolution kernel, the sixth layer is a 2×2 pooling layer, the seventh and eighth layers are convolutional layers consisting of two 3×3 convolution kernels, the ninth layer is a 2×2 pooling layer, the tenth and eleventh layers are convolutional layers with 3×3 convolution kernels, the twelfth layer is a standardized BN layer, the thirteenth layer is a convolutional layer with a convolution kernel size of 3×3, the fourteenth layer is a fully connected layer with Relu as the activation function, and the fifteenth layer is a fully connected layer with tanh as the activation function.

5. The method for designing broadband metamaterial absorbing structures based on deep learning according to claim 1, characterized in that: Step 5: Set the loss function and optimizer, train the forward prediction network and reversely design the deep convolutional neural network, where: In the process of training the forward prediction network, the loss function is defined as the mean square error MSE, and the Adam optimizer is used for training. The loss function of the trained network is stable at 10 -5 about.

6. The method for designing broadband metamaterial absorbing structures based on deep learning according to claim 1, characterized in that: Step 5: Set the loss function and optimizer, train the forward prediction network model and reversely design the deep convolutional neural network, where: In the process of training the reverse-designed deep convolutional neural network, the reverse-designed deep convolutional neural network is cascaded to the forward prediction network, and the forward prediction network is connected to the output of the reverse network. By minimizing the error between the input of the reverse network and the output of the forward prediction network, the loss function uses the mean square error MSE, and finally the loss function is stabilized at 10 -4 about.

7. A broadband metamaterial absorbing structure, characterized in that: A broadband metamaterial absorbing structure design method based on deep learning as described in any one of claims 1-6 is constructed.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for designing a broadband metamaterial absorbing structure based on deep learning according to any one of claims 1 to 6 is implemented to achieve the design of a broadband metamaterial absorbing structure.

9. A computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the method for designing a broadband metamaterial absorbing structure based on deep learning according to any one of claims 1 to 6 is implemented to achieve the design of a broadband metamaterial absorbing structure.

Images