Inverse Design Method of Strong Circular Dichroism Tunable Metasurface Based on Convolutional Neural Network

Abstract

The present invention provides an inverse design method for a strong circular dichroism tunable metasurface based on a convolutional neural network, including: S1, determining the basic structural unit of the metasurface structure and parameterizing the pattern etching region as a binary matrix; S2, randomly generating a number of etching pattern matrices; S3, performing simulation calculations on the metasurface structure in CST software to obtain a transmission spectrum line; S4, corresponding the etching pattern matrices randomly generated in S2 with the transmission spectrum lines obtained from the simulation calculations in S3 one by one, screening and organizing them into a sample set; S5, constructing a neural network model and training the neural network model with the sample set; S6, randomly generating a number of new etching pattern matrices again, inputting them into the trained neural network model, and obtaining the predicted transmission spectrum lines as output; S7, screening out the etching pattern matrices with circular dichroism higher than the first threshold. The present invention applies deep learning technology to the field of metasurface design, greatly improving the design efficiency of the metasurface structure.

Description

Technical Field

[0001] The present invention relates to the fields of deep learning and metasurface design, and particularly to an inverse design method for a strongly circular dichroism tunable metasurface based on a convolutional neural network. Background Art

[0002] Metamaterials are artificial composite materials composed of periodically arranged unit structures with sub-wavelength dimensions. They can achieve many physical responses by artificially designing the size and shape of the unit structures, especially unique functions that many natural materials do not possess. A metasurface can be regarded as the two-dimensional counterpart of a metamaterial, and its overall thickness is usually only a fraction or even a tenth of the working wavelength. Metasurfaces have the advantages of light weight, small volume, low loss, and easy integration, and can achieve flexible and effective control of the phase, amplitude, and polarization of electromagnetic waves.

[0003] Polarization selectivity means that when light passes through a metasurface, the metasurface can selectively reflect, absorb, or transmit light of a specific polarization state. Linear dichroism (LD) and circular dichroism (CD) are usually used to measure the effects of linear polarization and circular polarization selectivity of metasurfaces. This property is generally achieved by breaking the symmetry of the structure. These asymmetric structures will cause light waves of different polarization states to excite different electromagnetic modes on the metasurface, thereby generating different optical responses. How to obtain a metasurface structure that can generate strong circular dichroism is a key topic in the field of metasurface design. The current metasurface structure design relies on numerical simulation methods such as the finite-difference time-domain method and the finite element method. These methods largely rely on the experience of previous design templates and require a large amount of human participation and supervision. In addition, due to the limitations of simulation power and time, these numerical simulation methods can only adjust a limited number of design parameters when searching for the optimal structure, which greatly reduces the design efficiency.

[0004] In recent years, with the development of deep learning technology, researchers have begun to try to apply deep learning to solve metasurface design problems. However, most of the existing metasurface structure design methods based on deep learning technology can only be used to achieve some relatively simple design tasks for a single optical property and cannot design for the circular dichroism characteristics of metasurface structures. At the same time, the existing design processes and design algorithms generally have problems of low efficiency and insufficient diversity of design schemes, and need to be improved and perfected. Summary of the Invention

[0005] The purpose of the present invention is to provide an inverse design method for a strongly circular dichroism tunable metasurface based on a convolutional neural network to improve the design efficiency of metasurface structures in view of the defects existing in the prior art.

[0006] To achieve the above purpose, the present invention adopts the following technical solutions:

[0007] An inverse design method for a strong circular dichroism tunable metasurface based on a convolutional neural network, comprising the following steps:

[0008] S1. Determine the basic structural unit of the metasurface structure, and parameterize the pattern etching area in the basic structural unit as a binary matrix of a specific size;

[0009] S2. Based on the binary matrix of the specific size in S1, use Python to randomly generate a number of etching pattern matrices;

[0010] S3. For each etching pattern matrix generated in S2, model and generate the corresponding metasurface structure in CST software, and use CST software for simulation calculation to obtain the corresponding transmission spectrum line; the transmission spectrum line includes a left-handed circular polarization transmission spectrum line and a right-handed circular polarization transmission spectrum line;

[0011] S4. One-to-one correspond the randomly generated etching pattern matrices in S2 with the transmission spectrum lines obtained from the simulation calculation in S3, and screen and organize them into a sample set;

[0012] S5. Construct a neural network model with the etching pattern matrix as the input and the transmission spectrum line as the output, and use the sample set in S4 to train the neural network model;

[0013] S6. Using the same method as in S2, randomly generate a number of new etching pattern matrices again, input them into the trained neural network model, and obtain the predicted transmission spectrum line of the output;

[0014] S7. Screen according to the predicted transmission spectrum line output in S6, and select the etching pattern matrices with circular dichroism higher than the first threshold from the new etching pattern matrices generated in S6.

[0015] Further, in S1, the unit period of the basic structural unit of the metasurface structure p is 1000 nm, including a GST layer and a SiO 2 layer, the GST layer is stacked on the upper surface of the SiO 2 layer, and the pattern etching area is arranged in the middle of the GST layer.

[0016] Further, in S1, the method of parameterizing the pattern etching area in the basic structural unit as a binary matrix of a specific size is: after reserving a boundary of s = 20 nm at the edge of the GST layer with a unit period p of 1000 nm, the remaining area in the middle is used as the pattern etching area; the pattern etching area is divided into 48×48 side lengths wA square grid with a side length of 20 nm, corresponding to a 48×48 binary matrix; the etching pattern in the pattern etching area is represented by the binary matrix, where 0 represents that the square grid area is not etched, and 1 represents that the square grid area is etched.

[0017] Further, the step S2 specifically includes the following steps:

[0018] S201: Generate a 48×48 all-0 matrix, representing the initial state of the pattern etching area;

[0019] S202: Define the coordinates of the upper left corner of the all-0 matrix as the starting point (0, 0), and randomly generate i (3 ≤ i ≤ 8) coordinates ( x i , y i ), where 0 ≤ x i ≤ 48, 0 ≤ y i ≤ 48;

[0020] S203: Corresponding to the i coordinates ( x i , y i ), randomly generate i all-1 matrices with sizes of r i × c i respectively through a Python program, where r i is the number of rows of the all-1 matrix, c i is the number of columns of the all-1 matrix, 0 ≤ x i + c i ≤ 48, 0 ≤ y i + r i ≤ 48;

[0021] S204: Using the corresponding coordinates ( x i , y i ) as the starting point, write the i all-1 matrices into the all-0 matrix representing the pattern etching area according to the corresponding positions, and generate an etching pattern matrix representing the etching pattern.

[0022] Further, in S3, when outputting the transmission spectrum line, in the target design band, 40 discrete data points are continuously collected for the left-handed circularly polarized transmission spectrum line, and 40 discrete data points are continuously collected for the right-handed circularly polarized transmission spectrum line, and then output in the vector form of (1, 80).

[0023] Further, in S5, the constructed neural network model includes three convolutional layers, three pooling layers and seven fully connected layers; the three convolutional layers and the three pooling layers are cross-connected in series in turn, and a pooling layer is connected behind each convolutional layer; after passing through the three cross-connected convolutional layers and three pooling layers at the input end of the neural network model, it is then connected to the output end of the neural network model through seven fully connected layers; wherein, the convolutional kernel size of the convolutional layer is 2×2, the pooling type of the pooling layer is average pooling, and the number of neurons in each layer of the seven fully connected layers is 350, 300, 250, 200, 150, 100, 80 in turn.

[0024] Further, S7 includes the following steps:

[0025] S701. Using the circular dichroism of the predicted transmission spectrum line being higher than the second threshold as the preliminary screening condition, screening out the etching pattern matrix that meets the preliminary screening condition from the new etching pattern matrix generated in S6; the second threshold is less than the first threshold;

[0026] S702. Input the etching pattern matrix that meets the preliminary screening condition into the CST software, perform modeling and simulation in the CST software to obtain the actual transmission spectrum line of each etching pattern matrix; using the circular dichroism of the actual transmission spectrum line being higher than the first threshold as the re-screening condition, screening out the etching pattern matrix that meets the re-screening condition.

[0027] Further, the first threshold is 0.5 and the second threshold is 0.95.

[0028] Further, it also includes S8. Fabricating a metasurface structure according to the etching pattern matrix screened in S7; by changing the external temperature to change the phase of the GST layer, enabling it to switch between the crystalline state and the amorphous state, and realizing the dynamic regulation of the circular dichroism of the metasurface.

[0029] A reverse design method of a strongly circular dichroism tunable metasurface based on a convolutional neural network proposed by the present invention applies deep learning technology to the field of metasurface design, greatly improving the design efficiency of the metasurface structure. By digitizing the etching pattern of the metasurface structure into a binary matrix and further using a convolutional neural network to reversely design an etching pattern with strong circular dichroism, the design freedom of the metasurface structure is greatly improved.

[0030] The present invention uses the two-dimensional geometric pattern of a reconfigurable metasurface structure based on the phase change material GST as the input of a neural network model to predict the circularly polarized transmission spectrum in the amorphous state of GST, and then finds the optical structure pattern that can achieve tunable circular polarization selectivity (i.e., strong circular dichroism); the reverse design process adopted has the advantages of high degree of freedom, high design efficiency, strong feasibility, and easy implementation. The metasurface structure designed by the reverse design method of the present invention has the characteristics of strong circular dichroism and easy implementation.

[0031] The reverse design method proposed by the present invention overcomes the defect that the traditional metasurface structure design method is too cumbersome, simplifies the electromagnetic simulation work that requires a large amount of manual participation and the process of repeatedly adjusting model parameters, greatly improves the design efficiency and design freedom of the metasurface structure, and realizes a huge breakthrough in the ideas and methods of metasurface structure design. Brief Description of the Drawings

[0032] Figure 1 is the hierarchical structure diagram of the metasurface structure in the embodiment of the present invention.

[0033] Figure 2 is the schematic diagram of the method for parametrically representing the pattern etching area as a binary matrix in S1 of the embodiment of the present invention.

[0034] Figure 3 is the schematic diagram of the method for randomly generating an etching pattern matrix in S2 of the embodiment of the present invention.

[0035] Figure 4 is the schematic diagram of the structure of the neural network model in the embodiment of the present invention.

[0036] Figure 5 is the mean square error change diagram during the training process of the neural network model in the embodiment of the present invention.

[0037] Figure 6 is the etched pattern matrix with strong circular dichroism obtained after screening in S7 of the embodiment of the present invention.

[0038] Figure 7 is a metasurface structure obtained through the embodiment of the present invention.

[0039] Figure 8 is Figure 7 the circular dichroism curve of the metasurface structure shown under different phases of the GST layer. Detailed Embodiments

[0040] The technical solutions of the present invention will be described in detail below in conjunction with the drawings and specific embodiments.

[0041] A reverse design method of a strong circular dichroism tunable metasurface based on a convolutional neural network provided by an embodiment of the present invention includes the following steps:

[0042] S1. Determine the basic structural unit of the metasurface structure, and parameterize the pattern etching region in the basic structural unit as a binary matrix of a specific size;

[0043] S2. Based on the binary matrix of the specific size in S1, use Python to randomly generate a number of etching pattern matrices;

[0044] S3. For each etching pattern matrix generated in S2, model it in CST software to generate the corresponding metasurface structure, and use CST software for simulation calculation to obtain the corresponding transmission spectrum line; the transmission spectrum line includes a left-handed circular polarization transmission spectrum line and a right-handed circular polarization transmission spectrum line;

[0045] S4. One-to-one correspond the randomly generated etching pattern matrices in S2 with the transmission spectrum lines obtained from the simulation calculation in S3, and screen and organize them into a sample set;

[0046] S5. Construct a neural network model with the etching pattern matrix as the input and the transmission spectrum line as the output, and use the sample set in S4 to train the neural network model;

[0047] S6. Use the same method as in S2 to randomly generate a number of new etching pattern matrices again, input them into the trained neural network model, and obtain the predicted transmission spectrum line of the output;

[0048] S7. Screen according to the predicted transmission spectrum line output in S6, and select the etching pattern matrix with a circular dichroism higher than 0.95 from the new etching pattern matrices generated in S6.

[0049] Specifically, as Figure 1 shown, in S1, the unit period of the basic structural unit of the metasurface structure p is 1000 nm, including a GST (germanium antimony telluride) layer 1 and a SiO 2 layer 2, and the GST layer 1 is stacked on the upper surface of the SiO 2 layer 2, and the pattern etching region is arranged in the middle of the GST layer 1. Among them, the thickness of the GST layer 1 is 270 nm, and the SiO 2 layer 2 has a thickness of 450 nm, and the dielectric constant of the SiO 2 layer 2 is set to 2.1.

[0050] Further, as Figure 2 shown, in S1, the method of parameterizing the pattern etching region in the basic structural unit as a binary matrix of a specific size is: in the unit period pAfter reserving a boundary of s = 20 nm at the edge of the GST layer 1 with a length of 1000 nm, the remaining area in the middle is used as the pattern etching area; the pattern etching area is divided into 48×48 squares with a side length w of 20 nm, corresponding to a 48×48 binary matrix. In this way, the 2D etching pattern in the pattern etching area can be represented in a mathematical form through the binary matrix, where 0 represents that the square area is not etched, and 1 represents that the square area is etched.

[0051] After parameterizing the pattern etching area of the metasurface structure, a sample set for training the neural network model needs to be constructed. Each sample set data needs to contain both the first feature data as the input of the model and the second feature data as the output of the model. Among them, the first feature data is used to reflect the characteristics of the etching pattern, and the second feature data is used to reflect the circular dichroism of the metasurface structure adopting the etching pattern. In the embodiment of the present invention, a binary matrix corresponding to the size in S1 is used as the first feature data. At the same time, in the embodiment of the present invention, the numerical value of circular dichroism is not directly used as the second feature data, but the set of left-handed circular polarization transmission spectrum lines and right-handed circular polarization transmission spectrum lines is used as the second feature data. The main purpose is to facilitate the analysis and traceability of the simulation results and design results.

[0052] In the embodiment of the present invention, the acquisition process of the sample set data adopts a method of joint simulation of Python program and CST software, which reduces the difficulty of sample acquisition and improves the production efficiency of the sample set data. Among them, Python is responsible for randomly generating different etching patterns in a specific method, and the CST software is responsible for simulating and calculating the obtained metasurface structure to obtain its left-handed circular polarization transmission spectrum line and right-handed circular polarization transmission spectrum line.

[0053] Specifically, the S2 includes the following steps:

[0054] S201. Generate a 48×48 all-0 matrix, representing the initial state of the pattern etching area;

[0055] S202. Define the coordinates of the upper left corner of the all-0 matrix as the starting point (0, 0), and randomly generate i (3 ≤ i ≤ 8) coordinates ( x i , y i ), 0 ≤ x i ≤ 48, 0 ≤ y i ≤ 48;

[0056] S203. Corresponding to the i coordinates (x i , y i ) are randomly generated one - to - one through a Python program i full - one matrices with sizes of r i × c i , where r i is the number of rows of the full - one matrix, c i is the number of columns of the full - one matrix, 0 ≤ x i + c i ≤48, 0 ≤ y i + r i ≤48;

[0057] S204. Taking the corresponding coordinates ( x i , y i ) as the starting point, write i full - one matrices into the all - zero matrix representing the pattern etching area according to the corresponding positions, so as to generate an etching pattern matrix representing the etching pattern, which is used as the first feature data.

[0058] Taking the Figure 3 - shown etching pattern matrix as an example, each small square in the figure represents an 8×8 matrix, and the whole forms a pattern etching area with a size of 48×48. As Figure 3 shown, on the all - zero matrix with a size of 48×48, the Python program randomly generates 3 coordinates ( x 1 , y 1 ), ( x 2 , y 2 ) and ( x 3 , y 3 ), and at the same time generates 3 corresponding full - one matrices to form an etching pattern matrix. Among them, the size of the first full - one matrix A1 starting from the first coordinate ( x 1 , y 1 ) is 48×16, and the size of the second full - one matrix starting from the second coordinate ( x 2 , y 2) The size of the second all-1 matrix A2 starting from the third coordinate ( x 3 , y 3 ) is 16×16. The size of the third all-1 matrix A3 starting from the third coordinate (

[0059] ) is 16×16. After writing the three all-1 matrices into the corresponding positions of a 48×48 all-0 matrix, an etched pattern matrix with an overall S shape is formed.

[0060] Further, in S3, when outputting the transmission spectrum line, in the target design band of 182 - 184 THz, 40 discrete data points are continuously collected for the left-handed circularly polarized transmission spectrum line, and 40 discrete data points are continuously collected for the right-handed circularly polarized transmission spectrum line, and then output in the vector form of (1, 80) as the second feature data.

[0061] After collecting 30,000 sample data through the joint simulation of the Python program and CST software, appropriate screening and sorting are required to form a sample set. In S4 of this embodiment, data with more obvious circular dichroism (i.e., a larger difference between the left-handed circularly polarized transmission spectrum line and the right-handed circularly polarized transmission spectrum line) is selected from the 30,000 collected sample data as the initial data set. To ensure the diversity of the data, some etched patterns with circular dichroism of 0 can also be retained to enrich the diversity of the sample data. After screening, 10,000 sample data are retained and packaged into 10,000 sample sets.

[0061] After preparing the sample set through S2 to S4, a neural network model needs to be constructed. Since the etched patterns of the metasurface structure were processed into the form of a grayscale image matrix in the previous steps, and the convolutional neural network can quickly reduce the data dimension while extracting data features, which is very suitable for processing pictures.

[0062] As Figure 4 shown, in S5, the constructed neural network model includes three convolutional layers (CNN), three pooling layers (AvePooling), and seven fully connected layers; the three convolutional layers and the three pooling layers are cross-connected in series in turn, and a pooling layer is connected behind each convolutional layer; after passing through the cross-connected three convolutional layers and three pooling layers at the input end of the neural network model, it is then connected to the output end of the neural network model through seven fully connected layers; among them, the convolution kernel size of the convolutional layer is 2×2, the pooling type of the pooling layer is average pooling, and the number of neurons in each layer of the seven fully connected layers is 350, 300, 250, 200, 150, 100, 80 in turn. The learning rate of the neural network model is set to 7×10 -5 , the optimizer is selected as Adam, and the batch size is 32.

[0063] In the neural network model, after the single-channel 48×48 etched pattern matrix undergoes feature extraction through three convolutional layers and three pooling layers, a feature matrix with 16 channels and a size of 5×5 is finally obtained. Flattening this feature matrix yields a 1×400 vector, which is input into a seven-layer fully connected layer, and finally a 1×80 vector is obtained, corresponding to 80 data points (40 for left-handed circular polarization and 40 for right-handed circular polarization) of the left-handed circular polarization transmission spectrum line and the right-handed circular polarization transmission spectrum line.

[0064] After the neural network model is constructed, it is trained according to the following process: divide 10,000 sample sets, with 8,000 as the training set and the remaining 2,000 as the test set; use the training set to train the neural network model, compare the predicted transmission spectrum line output by the model with the actual transmission spectrum line obtained by CST software simulation in the training set, and calculate the error value. Repeatedly adjust the parameter weights of the neural network model according to the error until the mean square error (MSE) of the predicted transmission spectrum line converges significantly. As Figure 5 shown, after 25,000 trainings of the neural network model in this embodiment, it can be observed that the mean square error of the predicted transmission spectrum line output by the model converges significantly to around 0.005. Input the test set into the trained neural network model for testing and verification, and the average error of the test set is 0.13, which is lower than the set threshold and within the acceptable range, proving that the trained neural network model has high accuracy.

[0065] The trained neural network model has established a relatively accurate mapping relationship between the etched pattern matrix and the transmission spectrum line, and this neural network model can be used to assist in the design of the metasurface structure.

[0066] Specifically, in S6 of the embodiment of the present invention, 2,000 new etched pattern matrices are randomly regenerated using Python in the same method as in S2. These 2,000 new etched pattern matrices are respectively input into the trained neural network model to obtain the predicted transmission spectrum lines output.

[0067] After obtaining the predicted transmission spectrum lines output by the neural network model, the results need to be screened to obtain etched pattern matrices with high circular dichroism. Specifically, the predicted transmission spectrum lines with circular dichroism higher than 0.95 can be directly screened out through a program, and the corresponding etched pattern matrices are recorded.

[0068] As an improvement, to ensure the accuracy of the screening results, a method combining predicted transmission spectra and actual transmission spectra can be adopted for screening. At the same time, to improve the screening efficiency, computer programs can also be used for screening, which requires a specific execution standard and an instructable screening process. Specifically, in an improved embodiment of the present invention, in S7, the etched pattern matrix meeting the design requirements is screened out through the following steps:

[0069] S701. Taking the circular dichroism of the predicted transmission spectrum being higher than 0.5 as the primary screening condition, screen out the etched pattern matrix meeting the primary screening condition from the new etched pattern matrix generated in S6;

[0070] S702. Input the etched pattern matrix meeting the primary screening condition into CST software, perform modeling and simulation in CST software to obtain the actual transmission spectra of each etched pattern matrix; taking the circular dichroism of the actual transmission spectrum being higher than 0.95 as the secondary screening condition, screen out the etched pattern matrix meeting the secondary screening condition.

[0071] Since the predicted transmission spectrum in the embodiment of the present invention simultaneously includes 40 data points of the left-handed circularly polarized transmission spectrum and 40 data points of the right-handed circularly polarized transmission spectrum, in S701, only by setting the screening condition that the difference between the peak and valley of the predicted transmission spectrum is greater than 0.5, the predicted transmission spectrum with circular dichroism higher than 0.5 and its corresponding etched pattern matrix can be screened out. This screening condition has sufficient efficiency and executability and can achieve rapid screening after being combined with the instructions of the computer program. Further, combining the primary screening of S701 with the secondary screening of S702 can effectively balance the accuracy and efficiency of the screening results. As Figure 6 shown, after the primary screening of S701 and the secondary screening of S702, the embodiment of the present invention finally obtains 18 etched pattern matrices with strong circular dichroism.

[0072] As an improvement, the inverse design method of the embodiment of the present invention further includes S8. According to the etched pattern matrix screened out in S7, fabricate a metasurface structure; by changing the external temperature, change the phase of GST layer 1 to make it transform between the crystalline state and the amorphous state, so as to realize the dynamic regulation of the circular dichroism of the metasurface.

[0073] In this embodiment, an etched pattern matrix in the shape of a "cat's paw" is selected from the results screened out in S7. The etched pattern matrix in the shape of a "cat's paw" is composed of a combination of 6 rectangles with different sizes. Based on this etched pattern matrix, the fabricated metasurface structure is as Figure 7 shown.

[0074] As Figure 8As shown, after analyzing the circular polarization transmission spectrum of the metasurface structure, it is known that when the GST layer 1 is in the amorphous state, the circular dichroism curve of the metasurface structure has a peak of 0.97 at 182.9 THz. When the GST layer 1 changes from the amorphous state to the crystalline state under the influence of the external temperature, the value of circular dichroism can be changed from 0.97 to 0 at the maximum frequency (182.9 THz). Thus, by changing the external temperature, the phase of the GST layer 1 can be changed, enabling it to switch between the crystalline and amorphous states, and further enabling dynamic regulation of the metasurface. That is to say, the metasurface structure designed in this embodiment has strong circular dichroism that can be dynamically regulated.

[0075] The above-described embodiments merely represent several implementation manners of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the patent for the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent for the present invention shall be subject to the appended claims.

Claims

1. A convolutional neural network-based inverse design method for a metasurface with tunable strong circular dichroism, characterized in that: The following steps are involved: S1. Determine the basic structural unit of the metasurface structure, and parameterize the pattern etching area in the basic structural unit as a binary matrix of a specific size; S2, based on the binary matrix of a specific size in S1, using Python to randomly generate several etching pattern matrices; S3. For each etching pattern matrix generated in S2, a corresponding metasurface structure is modeled in the CST software, and a simulation calculation is performed using the CST software to obtain a corresponding transmission spectrum line; the transmission spectrum line includes a left-handed circularly polarized transmission spectrum line and a right-handed circularly polarized transmission spectrum line; S4, matching the randomly generated etching pattern matrix in S2 with the transmission spectrum obtained by simulation calculation in S3 one by one, and sorting them into a sample set after screening; S5, constructing a neural network model with the etching pattern matrix as input and the transmission spectrum as output, and training the neural network model using the sample set in S4; S6, using the same method as S2, randomly regenerate a number of new etching pattern matrices, input them into the trained neural network model, and obtain the output predicted transmission spectrum; S7, screening according to the predicted transmission spectrum output in S6, and screening out the etching pattern matrix having a circular dichroism higher than a first threshold from the new etching pattern matrix generated in S6; In S1, the unit period of the basic structural unit of the metasurface structure is p The thickness is 1000 nm, and the layer includes a GST layer and a SiO2 layer. The GST layer is stacked on the upper surface of the SiO2 layer, and the pattern etching area is arranged in the middle of the GST layer. In S1, the method of parameterizing the pattern etching area in the basic structural unit as a binary matrix of a specific size is as follows: p After reserving a boundary of s=20nm at the edge of the 1000nm GST layer, the remaining area in the middle is used as the pattern etching area; the pattern etching area is divided into 48×48 sides. w A square grid of 20 nm corresponds to a 48 × 48 binary matrix; The binary matrix is ​​used to represent the etching pattern in the pattern etching area, wherein 0 represents that the square area is not etched, and 1 represents that the square area is etched; In S5, the constructed neural network model includes three convolutional layers, three pooling layers and seven fully connected layers; the three convolutional layers and the three pooling layers are cross-connected in series, and each convolutional layer is connected to a pooling layer; the input end of the neural network model passes through the cross-connected three convolutional layers and the three pooling layers, and then passes through the seven fully connected layers to connect to the output end of the neural network model; wherein the convolution kernel size of the convolutional layer is 2×2, the pooling type of the pooling layer is average pooling, and the number of neurons in each layer of the seven fully connected layers is 350, 300, 250, 200, 150, 100, and 80, respectively.

2. The inverse design method of a strong circular dichroism tunable metasurface based on a convolutional neural network according to claim 1, characterized in that: The S2 specifically includes the following steps: S201, generating a 48×48 all-0 matrix to represent the initial state of the pattern etching area; S202, define the coordinates of the upper left corner of the all-0 matrix as the starting point (0, 0), and randomly generate it through the Python program i (3≤ i ≤8) coordinates ( x i , y i ), 0≤ x i ≤48,0≤ y i ≤48; S203, corresponding to the i Coordinates ( x i , y i ), randomly generated one by one through Python program i The sizes are r i × c i A matrix of all 1s, where r i is the number of rows of the matrix with all 1s, c i is the number of columns of the matrix with all 1s, 0≤ x i + c i ≤48,0≤ y i + r i ≤48; S204, with the corresponding coordinates ( x i , y i ) as the starting point, i The full-1 matrices are written into the full-0 matrix representing the pattern etching area according to the corresponding positions to generate an etching pattern matrix representing the etching pattern.

3. The inverse design method of a strong circular dichroism tunable metasurface based on a convolutional neural network according to claim 2, characterized in that: In S3, when outputting the transmission spectrum, in the target design band, 40 discrete data points are continuously collected for the left-handed circularly polarized transmission spectrum, and 40 discrete data points are continuously collected for the right-handed circularly polarized transmission spectrum, and then output in the form of a vector of (1, 80).

4. The inverse design method of a strong circular dichroism tunable metasurface based on a convolutional neural network according to claim 1, characterized in that: The S7 comprises the following steps: S701, taking the circular dichroism of the predicted transmission line higher than a second threshold as a preliminary screening condition, screening out an etching pattern matrix that meets the preliminary screening condition from the new etching pattern matrix generated in S6; the second threshold is less than the first threshold; S702. Input the etching pattern matrices that meet the initial screening conditions into the CST software, perform modeling and simulation in the CST software, and obtain the actual transmission spectra of each etching pattern matrix; use the circular dichroism of the actual transmission spectra being higher than the first threshold as the re-screening condition to screen out the etching pattern matrices that meet the re-screening conditions.

5. The inverse design method of a strong circular dichroism tunable metasurface based on a convolutional neural network according to claim 4, characterized in that: The first threshold is 0.5, and the second threshold is 0.

95.

6. The inverse design method of a strong circular dichroism tunable metasurface based on a convolutional neural network according to claim 1, characterized in that: It also includes S8, making a metasurface structure according to the etching pattern matrix selected by S7; changing the physical phase of the GST layer by changing the external temperature to make it switch between the crystalline state and the amorphous state, thereby realizing dynamic regulation of the circular dichroism of the metasurface.

Images