A deep neural network-based transparent microwave absorption device frequency response prediction and microstructure design method

Abstract

The application discloses a kind of transparent microwave absorption device frequency response prediction and microstructure design method based on deep neural network, belong to electromagnetic super surface design field.This method includes: according to transparent microwave absorption device microstructure pattern and its corresponding frequency response, array period, unit thickness construction data set.Data set is grouped into frequency response prediction model and is trained, and the frequency response prediction model of trained frequency response prediction model can replace electromagnetic simulation software and predict the frequency response of transparent microwave absorption device microstructure corresponding.Corresponding to the pattern of unit structure in data set part is input into variational autoencoder and is trained, and the variational autoencoder of trained variational autoencoder can generate pattern with the characteristics of data set.Covariance self-adapting evolutionary strategy is constructed by jointing variational autoencoder, frequency response prediction model and covariance self-adapting evolutionary strategy.The microstructure design method of transparent microwave absorption device is constructed.By optimizing latent vector in variational autoencoder, the transparent microwave absorption device microstructure with the frequency response of target frequency response can be directly obtained, and the design efficiency is greatly improved.Different transparent microwave absorption device microstructure can be obtained each time optimization, and the diversity of generated structure is improved, and the structure difficult to process is avoided to be designed.Further, the microstructure design method can design the transparent microwave absorption device microstructure through target ideal frequency response, and can meet different design requirements such as single peak, multiple peak and wideband.

Description

Technical Field

[0001] This invention relates to the field of electromagnetic metasurface design, and particularly to a method for predicting the frequency response and designing the microstructure of transparent microwave absorbing devices. Background Technology

[0002] Metasurfaces are two-dimensional metamaterials, artificial layered materials with a thickness less than the wavelength, forming a periodic array of structures on a subwavelength plane. Metasurfaces possess unique properties in the field of electromagnetic wave manipulation; by altering the microstructure pattern and structural parameters of the metasurface itself, it is possible to control the amplitude and phase of electromagnetic waves in specific wavelength bands. Currently, the field of metasurfaces has developed into an interdisciplinary area encompassing physics, informatics, and materials science. Metasurfaces can manipulate various characteristics of electromagnetic waves, with electromagnetic wave absorption being a crucial application. Current research on transparent microwave absorbing devices based on metasurfaces mainly focuses on increasing the absorption intensity, broadening the absorption bandwidth, and adjusting the absorption frequency. Regardless of the specific research direction, the design and optimization of the microstructure of transparent microwave absorbing devices are indispensable.

[0003] Traditional transparent microwave absorbing device microstructure design typically relies on researchers' years of design experience and extensive electromagnetic knowledge, continuously employing full-wave numerical simulation methods (such as the finite element method, finite-difference time-domain method, and finite integration technique (FIT)) to optimize the microstructure through trial and error. This consumes enormous time and computational resources. Furthermore, with the deepening research on absorbing metasurfaces, there is an urgent need for a simpler and faster design scheme for transparent microwave absorbing devices.

[0004] In recent years, deep learning technology has gradually entered the field of transparent microwave absorbing device design. This technology bypasses the extensive calculations of Maxwell's equations and employs a data-driven approach to design the microstructure of transparent microwave absorbing devices. In the field of using deep neural networks for transparent microwave absorbing device design, a priori dataset of transparent microwave absorbing devices is constructed based on frequency response requirements. This constructed dataset is then used to train the deep neural network. Training is the process of seeking the implicit relationship between input and output, namely the mapping relationship between frequency response, the microstructure pattern of the transparent microwave absorbing device, and structural parameters. The trained deep neural network model can extract implicit connections from the dataset that are difficult to derive from theoretical analysis and mathematical calculations. The extracted connections can be generalized by the deep neural network to data pairs outside the dataset; that is, when the relationship between data pairs remains unchanged, given the frequency response (microstructure pattern and structural parameters of the transparent microwave absorbing device), the deep neural network can provide the microstructure pattern and structural parameters (frequency response) of the transparent microwave absorbing device. Using deep neural networks for frequency response prediction or microstructure design of transparent microwave absorbing devices can shorten the time and reduce the difficulty.

[0005] In practical applications of transparent microwave absorbing devices, there are different target frequency response requirements, namely: the requirement for a single absorption resonant peak with high microwave absorption within the required bandwidth; the requirement for multiple absorption resonant peaks with freely designable microwave absorption frequency and absorption efficiency within the required bandwidth; and the requirement for broadband absorption with high microwave absorption in a wide band greater than 10 GHz. To quickly design the microstructure of a transparent microwave absorbing device that meets these target frequency responses, a deep neural network model is needed to accelerate the design process while ensuring accuracy.

[0006] 1. Patent 202310222586.5, "Optimization Method for Multi-Dimensional Multi-Channel Multiplexed Metasurface Holography Based on Reverse Design," describes a holographic metasurface optimization algorithm. This method can be applied to display, imaging, information storage, microscopy, and anti-counterfeiting encryption fields, but it provides a parameterized description of the metasurface, which is limited by the geometric parameters of the metasurface microstructure.

[0007] 2. Patent 202211341320.4, "An Electromagnetic Metamaterial Design Method Based on Deep Learning and Structural Variables," describes a metasurface design method based on a dual generator and the separation of the real and imaginary parts of the frequency response. This method constructs both a forward prediction model and an inverse design model. However, it involves a narrow bandwidth, only 8 GHz - 12 GHz, and fixes the structural parameters of the metasurface microstructure, resulting in poor design performance for various needs.

[0008] 3. Li Jiang et al. from Nanjing University of Posts and Telecommunications reported a neural network model based on a multilayer perceptron for designing nanostructured phase-manipulated metasurfaces. This network model accurately predicts phase values ​​by processing six geometric parameters and addresses the phase discontinuity problem in the reverse design of phase-manipulated metasurfaces by inserting a sigmoid function between the input layer and the hidden layers and using MAE as the loss function. However, this neural network model is only applicable to nanostructured phase-manipulated metasurfaces with six geometric parameters, exhibiting low generalization and portability. (Jiang L et al., “Neural network enabled metasurface design for phase manipulation”. Optics Express, 2021, 29(2):2521.)

[0009] 4. John A. Hodge et al. from Virginia Tech reported a deep neural network model based on a GAN variant, deep convolutional generative adversarial networks, to assist in the design of RF metasurfaces. They added a separate frequency response prediction neural network to the DC-GAN network model to quickly verify the frequency response of the synthesized metasurface microstructures. Although this model can only generate one specific metasurface microstructure for a specific target frequency response, if the result does not meet the fabrication and design requirements, then a trial-and-error approach needs to be used again for design. (John A. Hodge et al., “RF metasurface array design using deep convolutional generative adversarial networks”. 2019 IEEE International Symposium on Phased Array System & Technology (PAST). IEEE, 2019: 1-6)

[0010] In summary, the existing problems in the structural design and frequency response prediction of transparent microwave absorbing devices based on deep learning are as follows:

[0011] 1. Poor versatility of microstructure design models: They cannot simultaneously meet the ideal frequency response requirements of different applications within a wide frequency band, such as single-peak or multi-peak frequencies, adjustable amplitude, or broadband requirements. Generally, they can only perform reverse engineering design of structures for single-type frequency response requirements and frequency response within a narrow frequency band, resulting in poor versatility.

[0012] 2. The microstructure of transparent microwave absorbing devices obtained by microstructure design models is singular: Currently, most structural inverse models can only output a definite result for a single frequency response input. However, in datasets, there is often a one-to-many collision relationship between metasurface microstructures and electromagnetic responses, which leads to a decrease in design accuracy. Furthermore, when the output transparent microwave absorbing device microstructure has problems such as excessive fabrication difficulty or low light transmittance, only traditional design methods can be used to design a metasurface microstructure that meets the requirements.

[0013] 3. Poor versatility and low prediction accuracy of frequency response prediction models: Some studies fix the microstructure of transparent microwave absorbing devices to a specific shape and parameterize this shape, resulting in limited datasets. Furthermore, the architecture of the frequency response prediction models is relatively simple, and due to the influence of the dataset, they cannot achieve good fitting capabilities. Moreover, when predicting the frequency response of patterned microstructures, the array period and thickness are often fixed, further limiting their versatility. Summary of the Invention

[0014] This invention provides a method for frequency response prediction and microstructure design of transparent microwave absorbing devices based on deep neural networks, aiming to solve the above-mentioned problems. Its main components include:

[0015] A dataset of microstructures of transparent microwave absorbing devices was constructed, and a latent vector optimizer based on a variational autoencoder pattern generator, a frequency response prediction model, and a covariance adaptive adjustment evolution strategy was established.

[0016] The method for predicting the frequency response and designing the microstructure of a transparent microwave absorbing device based on a deep neural network is characterized in that the dataset is obtained in the following way:

[0017] The microstructure of the transparent microwave absorbing device was divided into 32×32 pixels. A pixel of 1 indicates that the area is covered by the transparent conductive material required by the microstructure, such as indium tin oxide or an ultrathin silver film, while 0 indicates that the area is not covered. The size of a single pixel is limited by the array period of the microstructure of the transparent microwave absorbing device, and is taken in the range of 1μm × 1μm to 300μm × 300μm. The microstructure pattern of the transparent microwave absorbing device was obtained by the "random rectangular rotational symmetry method", with a total of 35,000 groups. After setting the structural parameters, the microstructure of the transparent microwave absorbing device was simulated using electromagnetic simulation software to obtain its corresponding electromagnetic response.

[0018] The method for predicting the frequency response and designing the microstructure of a transparent microwave absorbing device based on a deep neural network is characterized by the following steps for constructing the microstructure pattern of the transparent microwave absorbing device using the "random rectangular rotational symmetry method":

[0019] First, generate an 8×16 pixel image with a grayscale value of 0. In this image, randomly generate 2-5 rectangles with a long side value ranging from 1-12 pixels and a narrow side value ranging from 1-6 pixels. Take the union of the areas covered by these rectangles and set its grayscale value to 1. Using this 8×16 pixel pattern as a base, mirror it along its y-axis to obtain a 16×16 pixel pattern. Then, rotate this pattern clockwise by 90°, 180°, and 270° in sequence. After that, stitch the original pattern with the rotated generated pattern to obtain a 32×32 pixel pattern. Finally, add a square with a side length ranging from 2-4 pixels in the center of the pattern.

[0020] The method for predicting the frequency response and designing the microstructure of a transparent microwave absorbing device based on a deep neural network is characterized in that the variational autoencoder pattern generator is divided into an encoder and a decoder. The encoder part takes the microstructure pattern of the transparent microwave absorbing device in the dataset as its overall input and outputs a latent vector. The decoder part takes the latent vector as its overall input and outputs the reconstructed microstructure pattern of the transparent microwave absorbing device as its overall output. Both the encoder and decoder parts use convolutional networks with residual connections and employ reparameterization techniques to connect them and obtain the gradient.

[0021] The overall workflow of the microstructure design method for transparent microwave absorbing devices is as follows:

[0022] I. Input the required target frequency response;

[0023] II. Initialize the latent vectors using a Gaussian distribution. The latent vectors are 20-dimensional, and the population size is k groups, where 50 ≤ k ≤ 100.

[0024] III. The initialized latent vectors are used to generate a transparent microwave absorption device microstructure pattern through the decoder part of the variational autoencoder, and the pattern is spliced ​​with the period and thickness mapped by the fully connected layer.

[0025] IV. Input the spliced ​​overall data into the frequency response prediction model to predict the frequency response;

[0026] V. Calculate the fitness function value for the jointly predicted frequency response and the required target frequency response;

[0027] VI. Select the 10 latent vectors with the largest fitness function values ​​as the basis for updating the sampling distribution parameters, and update the sampling distribution parameters accordingly.

[0028] VII. The preset number of training rounds is n, where 50 ≤ n ≤ 100. When the preset number of training rounds is reached or the absolute value of the fitness function is less than 1.00, the optimization stops and the optimal microstructure of the transparent microwave absorbing device corresponding to the latent vector is output. Multiple optimizations can yield outputs with similar frequency responses but different patterns of the transparent microwave absorbing device microstructure.

[0029] Furthermore, the method for predicting the frequency response and designing the microstructure of a transparent microwave absorbing device based on a deep neural network is characterized in that the dataset includes the microstructure pattern, period, thickness, and corresponding frequency response of the transparent microwave absorbing device.

[0030] Furthermore, the method for frequency response prediction and microstructure design of transparent microwave absorbing devices based on deep neural networks is characterized in that, during the training of the variational autoencoder, the algorithm model parameters are updated with reference to KL divergence and reconstruction loss. Specifically, the total loss, recall loss, and KL divergence are expressed as follows:

[0031]

[0032]

[0033]

[0034] In the formula This represents the KL divergence (regularization term). Indicates the losses from the recall. This represents the ratio between KL divergence and recall loss; Indicates mean square error. The standard deviation of the encoder output. This is the average value of the encoder output.

[0035] Furthermore, the method for frequency response prediction and microstructure design of transparent microwave absorbing devices based on deep neural networks is characterized in that the backbone of the frequency response prediction model is a convolutional network with residual connections, and a self-attention mechanism module and a recurrent neural network module are added; wherein, the self-attention mechanism module is used to obtain global information of the microstructure pattern of the transparent microwave absorbing device; the recurrent neural network module is used to obtain the relationship between square pixels and learn the physical properties of the microstructure of the transparent microwave absorbing device; the overall input of the frequency response prediction model is the stitched transparent microwave absorbing device microstructure pattern, period, and thickness, and the overall output is the predicted frequency response.

[0036] Furthermore, the method for frequency response prediction and microstructure design of transparent microwave absorbing devices based on deep neural networks is characterized in that the model parameters are updated during the training process of the frequency response prediction model, referring to the mean square error of the frequency response. Specifically, the mean square error is expressed as:

[0037]

[0038] In the formula, N is 501. This indicates the 501 S values ​​predicted by this deep neural network model. 11 The value of the parameter (frequency response), This represents the 501 simulated S values ​​in the dataset. 11 The value of the parameter (frequency response).

[0039] Furthermore, the method for predicting the frequency response and designing the microstructure of a transparent microwave absorbing device based on a deep neural network is characterized by defining an optimization objective function f for the covariance adaptive evolution strategy and the corresponding frequency response of the transparent microwave absorbing device microstructure. Fitness , is represented as:

[0040]

[0041] In the formula Indicates a given frequency response, This represents the frequency response of the microstructure of the transparent microwave absorbing device corresponding to the latent vector. This represents the weight between two objectives.

[0042] The innovativeness and good effects of this invention are:

[0043] 1. This invention proposes a method for constructing microstructure patterns for transparent microwave absorbing devices, namely the "random rectangular rotational symmetry method". The transparent microwave absorbing devices constructed by this method are characterized by being insensitive to electromagnetic wave polarization and having high average electromagnetic absorption efficiency.

[0044] 2. Based on convolutional neural networks, this invention proposes a frequency response prediction model for transparent microwave absorbing devices by combining self-attention mechanisms and recurrent neural networks. This model can predict the frequency response of transparent microwave absorbing devices with different structural parameters and microstructure patterns, and has high prediction accuracy and prediction speed far exceeding traditional electromagnetic simulation software.

[0045] 3. Based on variational autoencoders, frequency response prediction models, and evolutionary strategies for adaptive covariance adjustment, this invention proposes a microstructure design method for transparent microwave absorbing devices. This method can directly design the microstructure of transparent microwave absorbing devices according to the target ideal frequency response in different application scenarios within a wide frequency band. The designed microstructure has the required frequency response, and its design speed far exceeds that of traditional trial-and-error methods. Attached Figure Description

[0046] To more clearly illustrate the technical solution adopted in this invention, the following is a brief description of the accompanying drawings used in the technical methods adopted or proposed in this invention:

[0047] Figure 1 This is an overall structural diagram of the frequency response prediction model described in this invention.

[0048] Figure 2 This is a residual block structure diagram in the frequency response prediction model described in this invention.

[0049] Figure 3 This is a flowchart of the training process for the frequency response prediction model described in this invention.

[0050] Figure 4 This is a graph showing the change in loss values ​​of the frequency response prediction model described in this invention.

[0051] Figure 5 This is a prediction result diagram of the frequency response prediction model described in this invention.

[0052] Figure 6 This is an overall flowchart of the microstructure design method described in this invention.

[0053] Figure 7This is a diagram of the encoder structure in the variational autoencoder of the microstructure design method described in this invention.

[0054] Figure 8 This is a diagram of the decoder structure in the variational autoencoder of the microstructure design method described in this invention.

[0055] Figure 9 This is a basic block structure diagram of the encoder and decoder in the variational autoencoder of the microstructure design method described in this invention.

[0056] Figure 10 This is a flowchart of the latent vector optimization process based on the covariance adaptive adjustment evolution strategy and frequency response prediction model described in this invention.

[0057] Figure 11 This invention relates to the "random rectangle rotational symmetry method" pattern construction method.

[0058] Figure 12 This is the process for constructing a microstructure dataset for a transparent microwave absorbing device as described in this invention.

[0059] Figure 13 This is a display diagram of the microstructure pattern generated by the microstructure design method described in this invention.

[0060] Figure 14 This is a diagram showing the frequency response of the microstructure TE / TM generated by the microstructure design method described in this invention, and the frequency response of each microstructure.

[0061] Figure 15 This is a diagram showing the pattern and frequency response curve generated by the microstructure design method of the present invention for a single-peak target frequency response.

[0062] Figure 16 This is a diagram showing the pattern and frequency response curve generated by the microstructure design method of the present invention for multi-peak target frequency response.

[0063] Figure 17 This is a diagram showing the pattern and frequency response curve generated by the microstructure design method of the present invention for broadband target frequency response.

[0064] Figure 18 This is a comparison chart of the physical test curves and simulation curves of some microstructures designed by the microstructure design method described in this invention. Detailed Implementation

[0065] The frequency response prediction and microstructure design method proposed in this invention will be described in general below with reference to the accompanying drawings and a design example of a transparent microwave absorbing device. The overall requirements for implementing the embodiments are as follows:

[0066] 1. The proposed frequency response prediction model can predict the frequency response corresponding to the microstructure of transparent microwave absorbing devices with high accuracy.

[0067] 2. The proposed microstructure design method can design the microstructure of transparent microwave absorbing devices according to the different absorption characteristics required in the 5GHz-45GHz wide band. Specifically, for single-peak requirements, it can achieve an absorption efficiency of over 80.00% within the target bandwidth; for multi-peak requirements, it can design the peak absorption efficiency and absorption frequency while meeting the requirements for the number of absorption peaks; and for broadband requirements, it can achieve an absorption efficiency of no less than 80.00% within a bandwidth greater than 10GHz.

[0068] 3. Multiple different transparent microwave absorbing device microstructures can be designed for the same frequency response requirement, avoiding the reduction in design accuracy due to centralized one-to-many mapping of data, and eliminating microstructures with high processing difficulty and low light transmittance, thus selecting the optimal microstructure.

[0069] 4. The measured frequency response and the simulated frequency response of the designed transparent microwave absorbing device microstructure highly overlap in absorption frequency and amplitude, demonstrating consistency.

[0070] To address the aforementioned needs, this invention proposes a method for predicting the frequency response and designing the microstructure of a transparent microwave absorbing device. The overall structure diagram of the proposed frequency response prediction model is shown below. Figure 1 As shown, this is a combination of convolutional neural networks, recurrent neural networks, and self-attention mechanisms. Furthermore, the residual module in the frequency response prediction model is as follows... Figure 2 The diagram shows the residual connections of the convolution, with adaptive normalization replacing the usual batch normalization. The overall input to the frequency response prediction model is the stitched microstructure pattern, array period, and thickness, and the overall output is the predicted frequency response.

[0071] The objective function of the frequency response prediction model is the mean square error, expressed as:

[0072]

[0073] In the formula, N is 501. This indicates the 501 S predicted by this deep neural network model. 11 The value of the parameter (frequency response), This represents the 501 simulated S values ​​in the dataset. 11 The value of the parameter (frequency response). Figure 3 The diagram shows the overall training process for the frequency response prediction model. Figure 4 The image shows the change in loss value of the frequency response prediction model after 2000 training rounds, with the loss on the validation set as the benchmark; the lowest loss reached is 0.0009. After training, the frequency response prediction model can predict the frequency response of microstructures in transparent microwave absorbing devices, such as... Figure 5 The results of testing the trained frequency response prediction model on data pairs in the test set are shown, and it can be seen that the fit is relatively high.

[0074] The overall process of the proposed structural design method is as follows: Figure 6 As shown, it mainly consists of a pattern generator part of the variational autoencoder and a latent vector optimizer with a frequency response prediction model and a covariance adaptive adjustment evolution strategy. The variational autoencoder is further divided into an encoder and a decoder, as shown below. Figure 7 The image shows the overall structure of the encoder, whose input is a microstructure pattern from the dataset, and whose output is a 20-dimensional latent vector. Figure 8 The diagram shows the overall structure of the decoder, whose input is a 20-dimensional latent vector and whose output is the newly generated pattern after reconstruction. The basic block structure of the common module for both the encoder and decoder is shown below. Figure 9 As shown, the convolution consists of residual connections and upsampling / downsampling modules. The training objective function for the variational autoencoder is a weighted sum of KL divergence and mean squared error, with a weight of 0.005, expressed as:

[0075]

[0076]

[0077]

[0078] In the formula This represents the KL divergence (regularization term). Indicates the losses from the recall. This represents the ratio between KL divergence and recall loss. Indicates mean square error. The standard deviation of the encoder output. This represents the mean of the encoder output. A trained variational autoencoder can characterize the continuous latent space corresponding to patterns in the dataset and can generate entirely new patterns.

[0079] The optimization process for latent vectors in the structural design method proposed in this invention is as follows: Figure 10 As shown, the latent vectors are first initialized with a Gaussian distribution and a dimension of 20, with a population size of 50 groups. Then, the decoder part of the variational autoencoder generates patterns corresponding to the latent vectors. The structural parameters and patterns are then concatenated and input into the frequency response prediction model. The frequency response predicted by the frequency response prediction model and the expected frequency response are compared through the objective function f. Fitness Calculate the score, select the optimal latent vector to update the sampling distribution parameters, resample 50 latent vectors from the new distribution, and perform iterative optimization. Optimization stops when the maximum number of optimization rounds is reached or the fitness function value falls below 1.0, and the microstructure corresponding to the optimal latent vector is output. The optimization objective function is f. Fitness Represented as:

[0080]

[0081] In the formula Indicates a given frequency response, This represents the frequency response of the metasurface microstructure corresponding to the latent vector. This represents the weight between two objectives.

[0082] This invention provides a method for constructing microstructure patterns for transparent microwave absorbing devices, such as... Figure 11 As shown, this illustrates the basic idea of ​​the "random rectangle rotation symmetry method," and employs the following... Figure 12 The methods shown are combined to construct a dataset of microstructures for transparent microwave absorbing devices. To cover a wide band of 5 GHz to 45 GHz, a total of 35,000 patterns were generated, combining two different array periods of 7.00 mm and 2.10 mm, with a total of 70,000 data pairs in the dataset.

[0083] Two types of tests were conducted on the constructed and trained structural design method: one-to-many tests and design accuracy tests targeting the ideal frequency response. For the one-to-many tests, the original pattern and the generated pattern were as follows: Figure 13 As shown, the frequency response of the generated pattern is compared with that of the original pattern. Figure 14 As shown, it can be seen that the generated patterns have different shapes, but they all have similar frequency responses and are insensitive to polarization.

[0084] For design accuracy testing targeting the ideal frequency response, there are three types of tests: single-peak demand, multi-peak demand, and broadband demand. The test results for single-peak demand are as follows: Figure 15 As shown, the average absorption rates within the bandwidth were 87.8%, 90.4%, and 87.4%, respectively, meeting the design requirements. The test results for multi-peak demand are as follows... Figure 16 As shown, all tests meet the requirements for the number of absorption peaks, and the peak absorption frequency and absorption efficiency can be designed arbitrarily. The test results for broadband requirements are as follows... Figure 17 As shown, its average absorption in the 25GHz-40GHz frequency band is 87.7%. All three types of tests meet the frequency response requirements.

[0085] Three sets of transparent microwave absorbing device samples were fabricated, and their tested frequency response and simulated frequency response curves were compared as follows: Figure 18 As shown, its fit is relatively high.

[0086] In summary, the proposed frequency response prediction model exhibits high accuracy and can replace electromagnetic simulation software in the design of transparent microwave absorbing devices. The proposed microstructure design method can design microstructures for transparent microwave absorbing devices with different absorption characteristic requirements across a wide band of 5GHz-45GHz, and overcomes the one-to-many mapping problem, achieving a design speed far exceeding traditional design methods.

Claims

1. A method for predicting the frequency response and designing the microstructure of a transparent microwave absorbing device based on a deep neural network, characterized in that, The method includes: A dataset of microstructures of transparent microwave absorbing devices was constructed, and a latent vector optimizer based on a variational autoencoder pattern generator, a frequency response prediction model, and a covariance adaptive adjustment evolution strategy was established. The dataset was obtained through the following methods: The microstructure of the transparent microwave absorbing device was divided into 32×32 pixels. A pixel size of 1 indicates that the area is covered by the transparent conductive material required by the microstructure, namely indium tin oxide and an ultrathin silver film, while 0 indicates that the area is not covered. The size of a single pixel is limited by the period of the microstructure array of the transparent microwave absorbing device, and is taken in the range of 1μm × 1μm to 300μm × 300μm. The microstructure pattern of the transparent microwave absorbing device was obtained by the "random rectangular rotational symmetry method", with a total of 35,000 groups. After setting the structural parameters, the microstructure of the transparent microwave absorbing device was simulated using electromagnetic simulation software to obtain its corresponding electromagnetic response. The steps for constructing the microstructure pattern of the transparent microwave absorbing device using the "random rectangular rotational symmetry method" are as follows: First, generate an 8×16 pixel image with a grayscale value of 0. In this image, randomly generate 2-5 rectangles with a long side value ranging from 1-12 pixels and a narrow side value ranging from 1-6 pixels. Take the union of the areas covered by these rectangles and set its grayscale value to 1. Using this 8×16 pixel pattern as a base, mirror it along its y-axis to obtain a 16×16 pixel pattern. Then, rotate this pattern clockwise by 90°, 180°, and 270° in sequence. After that, stitch the original pattern with the rotated generated pattern to obtain a 32×32 pixel pattern. Finally, add a square with a side length ranging from 2-4 pixels in the center of the pattern. The variational autoencoder pattern generator is divided into two parts: an encoder and a decoder. The encoder part takes as input the microstructure pattern of the transparent microwave absorbing device in the dataset and outputs as a latent vector. The decoder part takes as input the latent vector and outputs as the reconstructed microstructure pattern of the transparent microwave absorbing device. Both the encoder and decoder parts use convolutional networks with residual connections and employ reparameterization techniques to connect them and obtain the gradient. The overall workflow of the microstructure design method for the transparent microwave absorbing device is as follows: I. Input the required target frequency response; II. Initialize the latent vectors using a Gaussian distribution. The latent vectors are 20-dimensional, and the population size is k groups, where 50 ≤ k ≤ 100. III. The initialized latent vectors are used to generate a transparent microwave absorption device microstructure pattern through the decoder part of the variational autoencoder, and the pattern is spliced ​​with the period and thickness mapped by the fully connected layer. IV. Input the spliced ​​overall data into the frequency response prediction model to predict the frequency response; V. Calculate the fitness function value for the jointly predicted frequency response and the required target frequency response; VI. Select the 10 latent vectors with the largest fitness function values ​​as the basis for updating the sampling distribution parameters, and update the sampling distribution parameters accordingly. VII. The preset number of training rounds is n, where 50 ≤ n ≤ 100. When the preset number of training rounds is reached or the absolute value of the fitness function is less than 1.00, the optimization stops and the optimal microstructure of the transparent microwave absorbing device corresponding to the latent vector is output. Multiple optimizations can yield outputs with similar frequency responses but different patterns of the transparent microwave absorbing device microstructure.

2. The method for frequency response prediction and microstructure design of a transparent microwave absorbing device based on a deep neural network as described in claim 1, characterized in that, The dataset includes the microstructure pattern, period, thickness, and frequency response of the transparent microwave absorbing device.

3. The method for frequency response prediction and microstructure design of a transparent microwave absorbing device based on a deep neural network as described in claim 1, characterized in that, During the training of the variational autoencoder, the algorithm model parameters are updated with reference to KL divergence and reconstruction loss. Specifically, the total loss, recall loss, and KL divergence are expressed as follows: In the formula This represents the KL divergence regularization term. Indicates the losses from the recall. This represents the ratio between KL divergence and recall loss; Indicates mean square error. The standard deviation of the encoder output. This is the average value of the encoder output.

4. The method for frequency response prediction and microstructure design of a transparent microwave absorbing device based on a deep neural network as described in claim 1, characterized in that, The frequency response prediction model has a convolutional network with residual connections as its backbone, and adds a self-attention mechanism module and a recurrent neural network module. The self-attention mechanism module is used to obtain global information about the microstructure pattern of the transparent microwave absorbing device. The recurrent neural network module is used to obtain the relationship between the square pixels and learn the physical properties of the microstructure of the transparent microwave absorbing device. The overall input of the frequency response prediction model is the stitched microstructure pattern, period, and thickness of the transparent microwave absorbing device, and the overall output is the predicted frequency response.

5. The method for frequency response prediction and microstructure design of a transparent microwave absorbing device based on a deep neural network as described in claim 1, characterized in that, During the training of the frequency response prediction model, the mean square error of the frequency response is used to update the model parameters. Specifically, the mean square error is expressed as: In the formula, N is 501. This indicates the 501 S predicted by this deep neural network model. 11 The value of the frequency response parameter. This represents the 501 simulated S values ​​in the dataset. 11 The value of the frequency response parameter.

6. The method for frequency response prediction and microstructure design of a transparent microwave absorbing device based on a deep neural network as described in claim 1, characterized in that, An optimization objective function f is defined for the covariance adaptive evolution strategy and the corresponding frequency response of the transparent microwave absorbing device microstructure. Fitness , is represented as: In the formula Indicates a given frequency response, This represents the frequency response of the microstructure of the transparent microwave absorbing device corresponding to the latent vector. This represents the weight between two objectives.

Images