Deep learning network model for reverse design of reflective metal metasurface unit

Abstract

The invention discloses a deep learning model for reverse design of a reflective metal metasurface unit, which comprises a VGG-based metasurface spectrum prediction network model and a C-DCGAN-based metasurface unit reverse design network model, and realizes a metasurface holographic imaging experiment under the frequency of 7GHz by using a single-bit coding metasurface of model reverse design. The trained predictor network model can accurately predict the reflection spectrum in the frequency band of 4 GHz-7 GHZ according to the input metasurface unit pattern, and the mean square error of the prediction result of the real part and the imaginary part of the reflection spectrum can reach 0.002 or below. The metasurface unit reverse design model based on the C-DCGAN design can output the corresponding metasurface unit pattern according to the input reflection spectrum, and the fitting degree of the reflection spectrum of the metasurface unit designed reversely and the given spectrum is very high. In an experiment, a network model is utilized to reversely design a single-bit coding metasurface for holographic imaging, and a single-bit coding metasurface holographic imaging experiment under 7GHz is completed on an established metasurface holographic imaging experiment platform. According to the deep learning model for reverse design of the reflective metal metasurface unit, a complicated simulation calculation process can be omitted for the metasurface design process, and the work can be applied to various applications with automatic design requirements of the reflective metal metasurface unit.

Description

Technical Field

[0001] The present invention relates to the field of optics, and in particular to a deep learning network model for reverse design of a reflective metal metasurface unit. Technical Background

[0002] Metamaterials are materials that have excellent ability to control material parameters. They have attracted much attention in the past two decades and have been extensively studied in various fields. In the early days, research focused on the control of effective parameters of three-dimensional structures, but the high loss and complex production of three-dimensional anisotropic dielectrics obviously limited their further development and application. Therefore, people proposed to planarize metamaterials, and gradually formed the concept of metasurfaces. As a two-dimensional planar electromagnetic artificial material, metasurfaces have the advantages of easy processing, low loss, and ultra-low profile without losing control capabilities.

[0003] Holographic imaging is a three-dimensional imaging technology based on the principle of light interference. It records the light wave information of the imaging target object to achieve the purpose of reconstructing the image of the target object. This technology must rely on a series of optical devices to achieve the imaging function. As a small structure with excellent light wave control capabilities, metasurfaces can play a good role in the field of holographic imaging. Holographic imaging based on metasurfaces has the advantages of miniaturization, integration, simple manufacturing and low cost. In recent years, the rapid development of controllable metasurfaces can achieve richer holographic imaging functions.

[0004] While metasurface research is ongoing, artificial intelligence technology (AI) has also developed rapidly and has been applied in many fields. AI technology uses various machine learning and deep learning algorithms to allow machines to imitate human behavior to solve various intelligent problems. Its basic principle is to let computers learn from a large amount of collected data until they make correct decisions. Since there is a lot of simulation work in metasurface research, especially the modeling and simulation process in the metasurface design process, which is time-consuming and labor-intensive, researchers have begun to use deep learning technology to realize the design research of metasurface structure and function to reduce the workload and accelerate the design process. Through machine learning algorithms or deep learning algorithms, metasurfaces that meet various phase amplitude requirements can be designed, thereby realizing various metasurface applications.

[0005] Combining deep learning methods with metasurface holographic imaging technology can utilize deep learning models to eliminate complicated simulation processes, realize the prediction of the corresponding spectrum of metasurface units and the reverse design of metasurface units, which can greatly reduce the workload of researchers. Summary of the invention

[0006] The purpose of the present invention is to provide a deep learning network model for the reverse design of reflective metal metasurface units, which can predict the corresponding spectrum of the metasurface unit and reverse design the metasurface unit.

[0007] In order to achieve the above object, the technical solution proposed by the present invention is as follows:

[0008] A convolutional neural network model for predicting the reflection spectrum of a reflective metal metasurface unit, characterized in that the reflection spectrum prediction network model is composed of a convolutional neural network with multiple convolutional layers and a small convolutional kernel based on a VGG model;

[0009] The convolutional neural network consists of 8 convolutional layers, 4 maximum pooling layers, 10 batch normalization layers, and 3 fully connected layers. The input data size is (32, 32, 1) and the output data size is (1, 1, 26).

[0010] The maximum pooling layer is located after the second, fourth, sixth, and eighth convolutional layers respectively to perform the maximum pooling operation;

[0011] The batch normalization layer is located after each layer in the network except the last layer;

[0012] The fully connected layer is located in the last three layers of the network;

[0013] The learning rate of the convolutional neural network model is 10e -4 , batchsize is 256, optimizer is Adam optimizer, activation function is Relu function, loss function is MSE;

[0014] The convolutional neural network training data set is 30,000 reflective metal metasurface unit pattern data and the reflectance spectrum corresponding to the metasurface unit in the 4GHz-7GHz frequency band, which is divided into a training set, a test set, and a validation set with sizes of 25,000, 4,000, and 1,000 respectively;

[0015] The reflective metal metasurface three-dimensional structure in the data set includes a dielectric layer with a thickness of 1.52 mm, a substrate layer with a thickness of 0.035 mm, a surface metal layer with a thickness of 0.035 mm, and a unit structure period of 10 mm;

[0016] The dielectric material used for the dielectric layer is Teflon material with a dielectric constant of 2.65 and a loss tangent value of 0.0019.

[0017] A generative adversarial network model for reverse design of reflective metal metasurfaces, characterized in that the generative adversarial network model combines a deep convolutional generative adversarial network with a conditional generative adversarial network to construct a C-DCGAN network model;

[0018] The generative adversarial network model consists of two models: a generator and a discriminator;

[0019] The first five layers of the generator model consist of a transposed convolution layer, a batch normalization layer, and a ReLu activation function layer, and the last layer consists of a transposed convolution layer and a Tanh activation function layer;

[0020] The generator model input size is (1,1,78), which consists of sampling noise (1,1,26), the real part of the reflection coefficient (1,1,26) and the imaginary part of the reflection coefficient (1,1,26), and the output size is (32,32,1);

[0021] The first six layers of the discriminator model are composed of convolutional layers, batch normalization layers, and LeakyReLu activation function layers. The last layer is a convolutional layer;

[0022] The discriminator model input size is (32,32,53), which is composed of a 1 / 4 hypersurface unit pattern (32,32,1) after binarization, a real part of the reflection coefficient (1,1,26) and an imaginary part of the reflection coefficient (1,1,26), and the output size is (1,1,1);

[0023] The learning rate of the generative adversarial network model is 10e -4 , batchsize is 128, optimizer is Adam optimizer, optimizer parameters β1 and β2 are 0.5 and 0.9 respectively, generator activation function is LeakyReLu, and discriminator activation function is ReLu;

[0024] The dataset used by the generative adversarial network model is 25,000 reflective metal metasurface unit pattern data and the reflectance spectrum corresponding to the metasurface unit in the 4GHz-7GHz frequency band, which is divided into a training set, a test set, and a validation set with sizes of 20,000, 4,000, and 1,000 respectively;

[0025] The three-dimensional structure and dielectric material parameters of the reflective metal metasurface unit in the data set are the same as the metasurface unit in the data set used by the convolutional neural network model.

[0026] The technical effects of the present invention are:

[0027] (1) The predictor network model proposed in the present invention can accurately predict the reflection spectrum of the input metasurface unit pattern in the 4GHz-7GHZ frequency band, and the mean square error of the prediction results of the real and imaginary parts of the reflection spectrum can reach less than 0.002.

[0028] (2) The inverse design model of the metasurface unit based on C-DCGAN design proposed in the present invention can output the corresponding metasurface unit pattern according to the input reflection spectrum. The reflection spectrum of the inversely designed metasurface unit has a high degree of fit with the given spectrum.

[0029] (3) In the present invention, a single-bit coded metasurface for holographic imaging was reversely designed using a network model, and a single-bit coded metasurface holographic imaging experiment at 7 GHz was completed on a constructed metasurface holographic imaging experimental platform. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The drawings in the specification, which constitute a part of this application, are used to provide a further understanding of the present invention. The illustrative embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0031] In the attached picture:

[0032] Figure 1 Schematic diagram of reflective metasurface holographic imaging. Metasurfaces are usually composed of arrayed microstructure units that can perform fine phase or amplitude modulation on incident light.

[0033] Figure 2 As a schematic diagram of the randomly generated metasurface pattern used in the training process, the code for generating random geometric patterns was first written based on Python. When generating random geometric patterns, circles, triangles, and rectangles are used as basic patterns, and a certain number of geometric patterns of random sizes are randomly distributed in the specified area of ​​the canvas to generate reflective metasurface unit patterns with practical preparation feasibility and randomness.

[0034] Figure 3 Schematic diagram of the supersurface unit structure.

[0035] Figure 4 Schematic diagram of the predictor network structure.

[0036] Figure 5 Schematic diagram of predictor loss.

[0037] Among them Figure 5 In the figure, the horizontal axis is the epoch number of the training process, and the vertical axis is the loss size of the model training process.

[0038] Figure 6 Compare the predictor prediction results with the simulation results.

[0039] Among them Figure 6 In the figure, the horizontal axis is the frequency (GHz) and the vertical axis is the value of the reflection coefficient.

[0040] Figure 7 Schematic diagram of the C-DCGAN network model structure.

[0041] Figure 8 Compare the actual input spectrum with the spectrum corresponding to the pattern generated by the generator.

[0042] Among them Figure 8 In FIG. 5 , the horizontal axis is the frequency (GHz) and the vertical axis is the value of the reflection coefficient.

[0043] Fig. 9 Calculate simulation results for the GS phase retrieval algorithm.

[0044] Fig.10 The metasurface unit pattern predicted using the network model.

[0045] Fig.11 These are the results of the metasurface electromagnetic simulation.

[0046] Fig.12 Picture of the actual preparation of the metasurface.

[0048] Fig.13 Schematic diagram of the microwave imaging experimental platform.

[0049] Fig.14 These are the metasurface holographic imaging results measured at different imaging distances. DETAILED DESCRIPTION

[0050] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with embodiments and drawings. The schematic implementation modes of the present invention and their description are only used to explain the present invention and are not intended to limit the present invention.

[0051] Figure 1 Schematic diagram of reflective metasurface holographic imaging.

[0052] Figure 2 Schematic diagram of randomly generated hypersurface patterns used in the training process.

[0053] Figure 3 Schematic diagram of the metasurface unit structure. The reflective metal metasurface three-dimensional structure includes a dielectric layer with a thickness of 1.52 mm, a substrate layer with a thickness of 0.035 mm, and a surface metal layer with a thickness of 0.035 mm. The unit structure period is 10 mm.

[0054] The dielectric material used for the dielectric layer is Teflon material with a dielectric constant of 2.65 and a loss tangent value of 0.0019.

[0055] Figure 4 This is a schematic diagram of the predictor network structure. The network consists of 8 convolutional layers, 4 maximum pooling layers, 10 batch normalization layers, and the last three layers of the network are fully connected layers.

[0056] Figure 5 This is a schematic diagram of the predictor loss. After 2000 epochs of learning optimization, the training error and test error of the real part dropped to 0.000258 and 0.00149, respectively, and the training error and test error of the imaginary part dropped to 0.000291 and 0.00168, respectively.

[0057] Figure 6 The prediction results of the predictor are compared with the simulation results. From the curve in the figure, it can be seen that the prediction results and the simulation results have a high degree of fit, which shows that the prediction neural network has sufficient reliability.

[0058] Figure 7 This is a schematic diagram of the C-DCGAN network model structure, which consists of two network models: the generator and the discriminator.

[0059] The first five layers of the generator model are composed of a transposed convolution layer, a batch normalization layer, and a ReLu activation function layer, and the last layer is composed of a transposed convolution layer and a Tanh activation function layer.

[0060] The first six layers of the discriminator model are composed of a convolutional layer, a batch normalization layer, and a LeakyReLu activation function layer. The last layer is a convolutional layer.

[0061] Figure 8 The actual input spectrum is compared with the spectrum corresponding to the pattern generated by the generator. It is obvious that the spectral response of the super-atomic unit structure generated by the generator for different target spectra is very close to the target spectral response, which proves the feasibility and reliability of the generator.

[0062] Fig. 9 To calculate the simulation results of the GS phase retrieval algorithm, we used the GS phase retrieval algorithm to calculate the phase distribution of the metasurface and the corresponding imaging pattern, and finally selected a frequency of 7 GHz and an imaging distance of 134 mm as the holographic imaging conditions.

[0063] Fig.10 In order to use the network model to predict the metasurface unit pattern, the two coding states of the metasurface, "1" and "0", correspond to the two reflection phases of -90° and 90°. For each phase, we use the network model to automatically generate four different metasurface units to form the entire metasurface, that is, two phases correspond to eight metasurface units, to fully verify the performance of the metasurface designed using the network model.

[0064] Fig.11To obtain the results of the electromagnetic simulation of the metasurface, after completing the design of the metasurface unit, the metasurface unit is modeled in CST, and the metasurface array is arranged according to the given metasurface phase distribution of the metasurface units with different patterns according to the GS algorithm. CST is used to simulate and obtain the electric field intensity distribution of the imaging plane 134mm away from the metasurface after the plane wave is incident on the metasurface and reflected.

[0065] Fig.12 This is a picture of the actual preparation of the metasurface. The PCB layout of the metasurface was drawn in AltiumDesigner and the metasurface was prepared and processed according to the PCB drawing.

[0066] Fig.13 This is a schematic diagram of the microwave imaging experimental platform. The experiment of metasurface microwave holographic imaging is carried out on an experimental platform composed of a vector network analyzer, a transmitting antenna, a receiving antenna, a computer, a stepper motor and a supporting motion mechanical device. The beam emitted by the transmitting antenna is incident on the metasurface and reflected, then received by the receiving antenna and the measured data is transmitted to the computer for processing. At the same time, the stepper motor drives the mechanical device to move according to the arrangement of the metasurface to complete the measurement of 32x32 units on the metasurface.

[0067] Fig.14 The holographic imaging results of the metasurface measured at different imaging distances show that the designed metasurface has the best imaging effect at an imaging distance of 134 mm. Compared with the simulation results, it can be seen that the actual imaging effect of the single-bit digital coding metasurface designed using the network model in the microwave band meets expectations.

[0068] In the present invention, a deep learning network model is designed for the reverse design of reflective metal metasurface units, which can predict the corresponding spectrum of the metasurface unit and reverse design the metasurface unit. The reflective metal metasurface reversely designed by the network model is used to realize microwave holographic imaging experiments at 7GHz.

[0069] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A convolutional neural network model for predicting the reflection spectrum of a reflective metal metasurface unit, characterized in that: The reflectance spectrum prediction network model is composed of a convolutional neural network with multiple convolutional layers and small convolutional kernels based on the VGG model; The convolutional neural network consists of 8 convolutional layers, 4 maximum pooling layers, 10 batch normalization layers, and 3 fully connected layers. The input data size is (32, 32, 1) and the output data size is (1, 1, 26). The maximum pooling layer is located after the second, fourth, sixth, and eighth convolutional layers respectively to perform the maximum pooling operation; The batch normalization layer is located after each layer in the network except the last layer; The fully connected layer is located in the last three layers of the network; The learning rate of the convolutional neural network model is 10e -4 , batchsize is 256, optimizer is Adam optimizer, activation function is Relu function, loss function is MSE; The convolutional neural network training data set is 30,000 reflective metal metasurface unit pattern data and the reflectance spectrum corresponding to the metasurface unit in the 4GHz-7GHz frequency band, which is divided into a training set, a test set, and a validation set with sizes of 25,000, 4,000, and 1,000 respectively; The reflective metal metasurface three-dimensional structure in the data set includes a dielectric layer with a thickness of 1.52 mm, a substrate layer with a thickness of 0.035 mm, and a surface metal layer with a thickness of 0.035 mm, and the unit structure period is 10 mm; The dielectric material used for the dielectric layer is Teflon material with a dielectric constant of 2.65 and a loss tangent value of 0.0019.

2. A generative adversarial network model for inverse design of reflective metal metasurfaces, characterized in that: The generative adversarial network model combines the deep convolutional generative adversarial network with the conditional generative adversarial network to construct a C-DCGAN network model; The generative adversarial network model consists of two models: a generator and a discriminator; The first five layers of the generator model consist of a transposed convolution layer, a batch normalization layer, and a ReLu activation function layer, and the last layer consists of a transposed convolution layer and a Tanh activation function layer; The generator model input size is (1,1,78), which consists of sampling noise (1,1,26), the real part of the reflection coefficient (1,1,26) and the imaginary part of the reflection coefficient (1,1,26), and the output size is (32,32,1); The first six layers of the discriminator model are composed of convolutional layers, batch normalization layers, and LeakyReLu activation function layers. The last layer is a convolutional layer; The discriminator model input size is (32,32,53), which is composed of a 1 / 4 hypersurface unit pattern (32,32,1) after binarization, a real part of the reflection coefficient (1,1,26) and an imaginary part of the reflection coefficient (1,1,26), and the output size is (1,1,1); The learning rate of the generative adversarial network model is 10e -4 , batchsize is 128, optimizer is Adam optimizer, optimizer parameters β1 and β2 are 0.5 and 0.9 respectively, generator activation function is LeakyReLu, and discriminator activation function is ReLu; The dataset used by the generative adversarial network model is 25,000 reflective metal metasurface unit pattern data and the reflectance spectrum corresponding to the metasurface unit in the 4GHz-7GHz frequency band, which is divided into a training set, a test set, and a validation set with sizes of 20,000, 4,000, and 1,000 respectively; The three-dimensional structure and dielectric material parameters of the reflective metal metasurface unit in the data set are as described in claim 1.

Images