CN110110398A - A Method for Automatic Metasurface Design Based on Convolutional Autoencoder - Google Patents

Abstract

The invention discloses a super-surface automatic design method based on a convolution self-encoder, which comprises the following steps: step 1, randomly generating a plurality of super-surface unit structures, and respectively calculating the reflection phase and amplitude of each super-surface by using electromagnetic simulation software; step 2, training a deep learning model by adopting a deep learning method based on a convolution self-encoder and simultaneously inputting the reflection phase and the amplitude obtained by calculation in the step 1 into the convolution self-encoder and outputting a corresponding super-surface unit structure; and 3, inputting the reflection phase and the amplitude of the design target into the deep learning model trained in the step 2, and completing feature extraction and matching between the features and the super-surface matrix by using a convolution self-encoder to obtain the super-surface structure required to be designed. The invention can simultaneously design the reflection amplitude and the reflection phase of the super surface, does not need a complex parameter scanning process, can automatically generate a surface structure, simplifies the design steps and has high design efficiency.

Description

A Method for Automatic Metasurface Design Based on Convolutional Autoencoder

technical field

The invention belongs to the technical field of metasurface design, and relates to an automatic metasurface design method based on a convolutional autoencoder.

Background technique

A metasurface is an artificial layered material with a thickness smaller than a wavelength. Through the design of the spatial distribution of phase mutations, metasurfaces can flexibly and effectively control the propagation characteristics of electromagnetic waves such as polarization, amplitude, phase, polarization mode, and propagation mode. A metasurface can be viewed as a two-dimensional array of planar subwave structures of metamaterials.

Currently widely used phase gradient metasurfaces, encoding metasurfaces, etc., need to simultaneously design the magnitude and phase of catadioptric refraction. However, most current methods for amplitude and phase control based on metasurfaces can only control a single parameter in the amplitude and phase, and it is difficult to achieve simultaneous control of the amplitude and phase. It is difficult for designers to simultaneously design the structure of the metasurface so that the amplitude and phase can reach the required size. At the same time, in the existing technology, designing a metasurface is not only time-consuming, but also computationally resource-intensive. It requires a complex parameter scanning process, modeling and parameter optimization, and the design steps are complicated. It is difficult for engineers to directly give the corresponding design goals. Metasurface, low design efficiency.

Contents of the invention

The purpose of the present invention is to provide a metasurface automatic design method based on a convolutional autoencoder, which solves the problems in the prior art that it is difficult to simultaneously regulate the metasurface amplitude and phase, and the design efficiency is low.

The technical scheme adopted in the present invention is, a kind of metasurface automatic design method based on convolution autoencoder, comprises the following steps:

Step 1: Randomly generate several metasurface unit structures, and use electromagnetic simulation software to calculate the reflection phase and amplitude of each metasurface;

Step 2, using the deep learning method based on the convolutional autoencoder, by inputting the reflection phase and amplitude calculated in step 1 into the convolutional autoencoder at the same time, and outputting the corresponding hypersurface unit structure, to train the deep learning model;

Step 3, input the reflection phase and amplitude of the design target into the deep learning model trained in step 2, use the convolutional self-encoder to complete the feature extraction and the matching between the feature and the hypersurface matrix, that is, to obtain the required design metasurface structure.

The present invention is also characterized in that:

The electromagnetic simulation software in step 1 is CST STUDIO SUITE MWS.

The specific process of step 1 is as follows:

First use the MATLAB software to output the matrix of the randomly generated metasurface unit structure, where the area marked "1" indicates that the area is filled with metal, and the area marked "0" indicates that the area is blank, and then use the electromagnetic simulation software CSTSTUDIO SUITE MWS Calculate the reflection phase and amplitude of the metasurface unit.

The specific process of training the deep learning model in step 2 is as follows:

Step 2.1, data acquisition generation and preprocessing process: use MATLAB language to draw the image of the reflection phase and amplitude of the metasurface unit, perform normalization operation on the grayscale and pixel features of the image, and set the image to 32×32 ×1 single-channel grayscale image, the pixels of the image are limited to between 0 and 1, and the image with pixels between 0 and 1 is used as the input of the deep learning model;

Step 2.2, the process of using convolutional autoencoder to extract features: input the image with pixels between 0 and 1 obtained in step 2.1 into the convolutional autoencoder, firstly through several convolutions, pooling operations and corresponding number of The upsampling and deconvolution operations compress and reconstruct the input reflection phase and amplitude curve images into a representation space, and then extract the data according to this representation space to obtain the final output data matrix; then the convolution self-encoder The encoding part of is disassembled, the input image data is encoded, the original input image is compressed into a representation vector, the convolution kernel is iteratively adjusted through the activation function I until the loss function I is minimized, and the representation vector when the loss function I is minimized is obtained;

Step 2.3, the metasurface structure matrix matching process based on the convolutional autoencoder: use the fully connected artificial neural network for training through the activation function II and the loss function II, and input the representation vector matrix generated by the encoding part of the convolutional autoencoder, Output the metasurface structure matrix, until the loss function II is minimized, the matching between the representation vector features and the hypersurface structure when the loss function I is minimized in step 2.2 is completed, and the training of the deep learning model is completed.

The activation function I is Where x represents the output data of each layer of the neural network after linear weighting; the loss function I is the cross entropy cost function Where x represents the number of output data, y is the desired output, and y' is the actual output.

The activation function II adopts ReLU=max(0,x), where x represents the output data of each layer of the neural network after the linear weighted summation; the loss function II is the mean square error Where yi is the desired output, and y′ _i is the actual output.

The beneficial effect of the present invention is: it solves the problems in the prior art that it is difficult to control the amplitude and phase of the metasurface at the same time, and the design efficiency is low, and the metasurface structure matrix Matching its electromagnetic properties can realize the design of metasurface reflection phase and reflection amplitude at the same time; and the method of using convolutional autoencoder can realize unsupervised learning. This method imitates the construction of biological visual attention mechanism, combined with convolutional neural network Operations such as convolution and pooling of the network do not require complex parameter scanning processes, modeling and parameter optimization, and the design of metasurfaces can be automated, which simplifies the design steps and has high design efficiency.

Description of drawings

Fig. 1 is a schematic diagram of the unit structure of the metasurface described in a metasurface automatic design method based on a convolutional autoencoder of the present invention;

Fig. 2 is the flow chart of deep learning model training based on convolution autoencoder in the present invention;

Fig. 3 is a flow chart of metasurface structure design based on convolutional autoencoder in the present invention;

Fig. 4 is a structural representation of the convolutional autoencoder described in a metasurface automatic design method based on the convolutional autoencoder of the present invention;

Fig. 5 (a) is the design goal of reflection magnitude and phase proposed in the embodiment of the present invention; Fig. 5 (b) is the value of metasurface reflection magnitude and phase designed for the design goal in Fig. 5 (a) .

In the figure, 1. Dielectric layer, 2. Structure layer, 3. Metal backplane.

Detailed ways

The present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

A kind of metasurface automatic design method based on convolution autoencoder of the present invention, as shown in Figure 1, 2 and 3, comprises the following steps:

Step 1: Randomly generate several metasurface unit structures, and use electromagnetic simulation software to calculate the reflection phase and amplitude of each metasurface;

Step 2, using the deep learning method based on the convolutional autoencoder, by inputting the reflection phase and amplitude calculated in step 1 into the convolutional autoencoder at the same time, and outputting the corresponding hypersurface unit structure, to train the deep learning model;

Step 3, input the reflection phase and amplitude of the design target into the deep learning model trained in step 2, use the convolutional self-encoder to complete the feature extraction and the matching between the feature and the hypersurface matrix, that is, to obtain the required design metasurface structure.

The electromagnetic simulation software in step 1 is CST STUDIO SUITE MWS.

The specific process of step 1 is as follows:

First use the MATLAB software to output the matrix of the randomly generated metasurface unit structure, where the area marked "1" indicates that the area is filled with metal, and the area marked "0" indicates that the area is blank, and then use the electromagnetic simulation software CSTSTUDIO SUITE MWS Calculate the reflection phase and amplitude of the metasurface unit.

The specific process of training the deep learning model in step 2 is as follows:

Step 2.1, data acquisition generation and preprocessing process: use MATLAB language to draw the image of the reflection phase and amplitude of the metasurface unit, and set the grayscale and pixels of the image, and set the image as a single unit of 32×32×1 Channel grayscale image, since the pixel range of the grayscale image is between 0 and 255, in order to make the feature distribution more uniform, the feature is normalized, the pixel of the image is limited to between 0 and 1, and the pixel is used to The image between 0 and 1 is used as the input of the deep learning model;

Step 2.2, the process of using convolutional autoencoder to extract features: input the image with pixels between 0 and 1 obtained in step 2.1 into the convolutional autoencoder, firstly through several convolutions, pooling operations and corresponding number of The upsampling and deconvolution operations compress and reconstruct the input reflection phase and amplitude curve images into a representation space, and then extract the data according to this representation space to obtain the final output data matrix; then the convolution self-encoder The encoding part of is disassembled, the input image data is encoded, the original input image is compressed into a representation vector, the convolution kernel is iteratively adjusted through the activation function I until the loss function I is minimized, and the representation vector when the loss function I is minimized is obtained;

Step 2.3, the metasurface structure matrix matching process based on the convolutional autoencoder: use the fully connected artificial neural network for training through the activation function II and the loss function II, and input the representation vector matrix generated by the encoding part of the convolutional autoencoder, Output the metasurface structure matrix, until the loss function II is minimized, the matching between the representation vector features and the hypersurface structure when the loss function I is minimized in step 2.2 is completed, and the training of the deep learning model is completed.

The activation function I is Where x represents the output data of each layer of the neural network after linear weighting; the loss function I is the cross entropy cost function Where x represents the number of output data, y is the desired output, and y' is the actual output.

The activation function II adopts ReLU=max(0,x), where x represents the output data of each layer of the neural network after the linear weighted summation; the loss function II is the mean square error Where yi is the desired output, and y′ _i is the actual output.

Example

As shown in Figure 1, the metasurface unit structure in step 1 is composed of a dielectric layer 1, a structural layer 2 and a metal backplane layer 3, the lower surface of the dielectric layer 1 is the metal backplane layer 3, and the upper surface of the dielectric layer 1 is Structural layer 2, the material of the dielectric layer 1 of the supersurface unit structure in the present embodiment is glass fiber epoxy resin copper-clad laminate (Fr4), the thickness of the dielectric layer 1 is 1.5mm, and the thickness of the metal backplane layer 3 is 0.018mm. The material of the structure layer 2 and the metal backplane layer 3 in the microwave frequency band is copper; the structure layer 2 is divided into 64 areas marked with "0" or "1", and the area marked with "1" means that the area is filled with metal , the area marked "0" indicates that the area is blank, so the structural layer 2 can be represented by a matrix, and the metasurface unit structure can also be represented by a matrix when the material and thickness of the dielectric layer 1 are fixed values.

As shown in Figures 2, 3 and 4, first use MATLAB software to output the 8*8 matrix of 2000 sets of metasurface unit structures randomly generated, and then use the electromagnetic simulation software CST STUDIO SUITE MWS to calculate the reflection phase of 2000 sets of metasurface units and magnitude.

Since the input two-dimensional image is a series of two-dimensional arrays from the computer point of view, it is intuitively difficult to extract appropriate features from it to match the output matrix. Therefore, the feature extraction operation is first performed on the input image. The unsupervised learning of images can be realized by using the method of convolutional self-encoder. This method imitates the construction of biological visual attention mechanism, and combines the convolution and pooling operations of convolutional neural network to realize feature extraction.

As shown in Figure 3, for a single-channel input image with a size of 32×32×1, convolution operation and pooling operation are performed on it. The size and depth of each convolution kernel are different, but the stride is 1, and the complement The zero method is equivalent padding (same padding), and the pooling operation uses the maximum pooling function (MaxPooling). The basic parameters of the model encoding end in Figure 3 are:

Encoding part:

The first layer of convolution: 32 7×7 convolution kernels, the step size is 1, and the padding method is equivalent padding (same padding);

The first layer of pooling: the pooling filter size is 4×4, and the maximum pooling function;

The second layer of convolution: 64 convolution kernels of 5×5, and the remaining parameters are consistent with the first layer of convolution;

The second layer of pooling: the pooling filter size is 4×4, and the maximum pooling function;

The third layer of convolution: 128 3×3 convolution kernels, and the rest of the parameters are consistent with the first layer of convolution;

The third layer of pooling: the pooling filter size is 2×3, and the maximum pooling function.

After the above operations are completed, the matrix with the original input size of 32×32×1 is compressed into a matrix of 1×1×128, and then the data is reconstructed through the decoding part. The basic parameters of the decoding end of the model in Figure 3 are:

The first layer of upsampling: the sampling kernel size is 2×2;

The first layer of deconvolution: 64 3×3 convolution kernels, the step size is 1, and the padding method is equivalent padding (same padding);

Upsampling on the second layer: the sampling kernel size is 4×4;

The second layer of deconvolution: 32 5×5 convolution kernels, the step size is 1, and the padding method is equivalent padding (same padding);

Upsampling on the third layer: the sampling kernel size is 4×4;

The second layer of deconvolution: 1 convolution kernel of 7×7, the step size is 1, and the padding method is equivalent padding (same padding).

After the above decoding operation is completed, the 1×1×128 matrix compressed by the encoding end is reconstructed into a 32×32×1 matrix.

The execution process of the convolutional autoencoder is the process of finding the minimum value of the cross-entropy cost function. According to the given input image, the convolution kernel is adjusted iteratively through the activation function I, so that the loss function I is minimized, and the reconstructed image is most similar to the original input image. After the convolutional self-encoder training is completed, its encoding part is disassembled, and the compressed 1 × 1 × 128 matrix of the encoding part is taken out as the result of feature extraction, that is, the feature dimension is compressed from the original 32 × 32 × 1 = 1024 To 1×1×128=128 dimensions, the feature dimension becomes 12.5% of the original, and has little effect on the effect of the image.

The hypersurface structure matrix matching process based on the convolutional autoencoder is to match the newly extracted representation vector features with the hypersurface structure matrix, using a fully connected artificial neural network for training, and inputting the encoding end of the convolutional autoencoder The resulting matrix of representation vectors outputs a matrix of metasurface structures.

In the embodiment of the present invention, in order to realize the matching between the newly extracted 128-dimensional feature and the metasurface structure matrix, the input of the network is the newly generated 128-dimensional feature matrix, and the activation function II is used to add nonlinear factors to the neural network, and the loss When function II is minimized, the output size is 8×8 metasurface structure matrix. The model can be defined as:

Y=f(Xβ)

Among them, X is the input vector, β is the parameter vector, Y is the output vector, and f(*) is the activation function.

The learning process of the neural network is the process of seeking the minimum value of the mean square error. According to the initial β parameter setting, an output y′ _i is obtained, and the value of the mean square error between y′ _i and the expected output y _i is calculated. The gradient descent algorithm determines the minimum value of the mean square error, and iteratively updates the setting of each β parameter in the neural network through the back propagation algorithm (BP algorithm), so as to obtain the value of the β parameter of the neural network under the condition of minimizing the loss function II.

After the model training is completed, as shown in Figure 5(a), the design target in this embodiment is a metasurface unit with wave-absorbing properties at 15.9GHz and 18GHz, and the phase has a 180-degree change from 8GHz to 14.5GHz .

Taking the design goals of phase and reflection coefficient as input, the deep learning model will automatically output an 8*8 matrix:

In order to verify the accuracy of the design results, the structure represented by the matrix is calculated in CST STUDIO SUITE MWS, as shown in Figure 5(b), the electromagnetic characteristics of the model are very close to the design goals, and there are two A strong absorption peak presents a strong wave-absorbing characteristic; from 8 GHz to 14.5 GHz, the metasurface achieves a 180-degree phase change, which proves the validity of the design results.

A metasurface automatic design method based on a convolutional self-encoder of the present invention has the advantage that the reflection amplitude and phase of the metasurface structure are connected through the metasurface design method based on a convolutional self-encoder. In the process, the reflection amplitude and reflection phase of the metasurface can be designed at the same time; the method of using the convolutional autoencoder can realize unsupervised learning without modeling and parameter optimization, which can automatically generate the metasurface structure, simplifying the The design steps improve the design efficiency.

Claims (6)

1. a kind of super surface automatic design method based on convolution self-encoding encoder, which comprises the following steps:

Step 1, several super surface cell structures are generated at random, calculate separately the anti-of each super surface using electromagnetic simulation software Penetrate phase and amplitude；

Step 2, using the deep learning method based on convolution self-encoding encoder, pass through the reflected phase that will be calculated in step 1 Convolution self-encoding encoder is inputted simultaneously with amplitude, exports corresponding super surface cell structure, Lai Xunlian deep learning model；

Step 3, it will train in the deep learning model finished, use in the reflected phase of design object and amplitude input step 2 The convolution self-encoding encoder completes the matching between feature extraction and feature and super surface matrix, that is, designs required for obtaining Super surface texture.

2. a kind of super surface automatic design method based on convolution self-encoding encoder according to claim 1, which is characterized in that The electromagnetic simulation software in the step 1 is CST STUDIO SUITE MWS.

3. a kind of super surface automatic design method based on convolution self-encoding encoder according to claim 2, which is characterized in that Detailed process is as follows for the step 1:

The matrix of the super surface cell structure generated at random using the output of MATLAB software first, wherein being labeled as the region of " 1 " It indicates that the area filling has metal, indicates the region blank labeled as the region of " 0 ", reuse the electromagnetic simulation software CST STUDIO SUITE MWS carries out reflected phase and amplitude that the super surface cell is calculated.

4. a kind of super surface automatic design method based on convolution self-encoding encoder according to claim 1, which is characterized in that Detailed process is as follows for training deep learning model in the step 2:

Step 2.1, data acquisition generate and preprocessing process: using MATLAB language draw super surface cell reflected phase and Operation is normalized in the image of amplitude, gray scale and pixel characteristic to the image, sets image as 32 × 32 × 1 single channel Grayscale image limits the pixel of image between 0~1, and using image of this pixel between 0~1 as deep learning model Input；

Step 2.2, using convolution self-encoding encoder extract feature process: by the pixel obtained in the step 2.1 0~1 it Between image input convolution self-encoding encoder, pass through convolution several times, pondization operation and the accordingly up-sampling of number, deconvolution first The reflected phase of input and amplitude curve compression of images are reconfigured in a representation space by operation, then further according to this expression The output data matrix that space extracts to the end to data；Then under the coded portion of convolution self-encoding encoder being disassembled Come, input image data is encoded, original input image is compressed into expression vector, is rolled up by I iteration adjustment of activation primitive Product core obtains expression vector when loss function I minimizes until the minimum of loss function I；

Step 2.3, the super surface texture matrix matching process based on convolution self-encoding encoder: pass through activation primitive II and loss function II is trained using the artificial neural network connected entirely, the expression moment of a vector that the coded portion of input convolution self-encoding encoder generates Battle array exports super surface texture matrix, until when loss function I minimizes in completion step 2.2 when loss function II minimizes The matching indicated between vector characteristics and super surface texture, deep learning model training finishes.

5. a kind of super surface automatic design method based on convolution self-encoding encoder according to claim 4, which is characterized in that The activation primitive I isWherein x represents the output data after each layer of linear weighted function of neural network； The loss function I is cross entropy cost functionWherein x represents output data Number, y be desired output, y' be actual output.

6. a kind of super surface automatic design method based on convolution self-encoding encoder according to claim 4, which is characterized in that The activation primitive II uses ReLU=max (0, x), and wherein x represents the output after each layer line weighted sum of neural network Data；The loss function II is mean square errorWherein y_iFor desired output, y '_iFor reality output.