Intelligent reverse design method and system for polarization conversion metasurface based on deep learning

Abstract

The present invention relates to the field of artificial intelligence technology, and specifically to a method and system for intelligent reverse design of polarization conversion metasurfaces based on deep learning, including using polarization degree angle and quasi-polarization degree angle characterization mechanism to characterize polarization conversion metasurfaces to obtain continuity characterization; constructing a data set model based on continuity characterization; completing deep learning network training based on the data set model, and using this to reverse design polarization conversion metasurfaces, thereby obtaining a metasurface design scheme. The present invention implements a polarization conversion metasurface functional characterization mechanism by using polarization degree angle and quasi-polarization degree angle to characterize polarization conversion metasurfaces, solving the problem of broadband phase difference discontinuity in traditional characterization methods and infinite characterization variables in axial ratio characterization methods.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for intelligent reverse design of a polarization conversion metasurface based on deep learning. Background Art

[0002] The rapid development of the field of artificial intelligence, represented by deep learning technologies such as convolutional neural networks and generative adversarial networks, has made the intelligent reverse design of high-performance polarization conversion metasurfaces in seconds possible.

[0003] When faced with the reverse design of polarization conversion metasurfaces based on deep learning technology, the traditional characterization method using the relationship between the amplitude and phase of the electric field components has the problem of broadband phase difference discontinuity, which brings great difficulties to the research on intelligent reverse design methods of polarization conversion metasurfaces based on deep learning. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for intelligent inverse design of polarization conversion metasurface based on deep learning, aiming to solve the problem of broadband phase difference discontinuity in traditional characterization methods.

[0005] To achieve the above objectives, in a first aspect, the present invention provides a method for intelligent inverse design of a polarization conversion metasurface based on deep learning, comprising the following steps:

[0006] The polarization conversion metasurface is characterized by using polarization angle and quasi-polarization angle characterization mechanism to obtain a continuous characterization.

[0007] constructing a data set model based on the continuity representation;

[0008] Based on the data set model, deep learning network training is completed, and the polarization conversion metasurface is reversely designed to obtain a metasurface design solution.

[0009] The specific method of using the polarization angle and quasi-polarization angle characterization mechanism to characterize the polarization conversion metasurface and obtain continuity characterization is as follows:

[0010] The polarization degree angle continuity is used to characterize the linear-circular polarization conversion metasurface, and the quasi-polarization degree angle continuity is used to characterize the cross-polarization conversion metasurface, thereby achieving the continuity characterization of the polarization conversion metasurface.

[0011] The polarization angle has a value range of -π / 4 to π / 4, and when the polarization angle is ±π / 4, it indicates that the wave is left-handed circularly polarized;

[0012] The value range of the quasi-polarization angle is 0 to π / 2, and when the quasi-polarization angle is π / 2, it indicates that the wave is cross-polarized; when the quasi-polarization angle is 0, it indicates that the cross-polarization conversion function is not realized.

[0013] The specific method of constructing the data set model based on the continuity representation is:

[0014] Electromagnetic simulation is used to obtain the metasurface propagation coefficient, and the relationship between the polarization angle and the quasi-polarization angle parameter response and the metasurface propagation coefficient is established to obtain a metasurface continuity characterization relationship;

[0015] Describing the unit structure arrangement of the polarization conversion metasurface based on the encoding matrix of the metasurface;

[0016] Iterating the arrangement of the unit structure using an iterative optimization algorithm based on the continuity representation relationship of the hypersurface to obtain an optimized coding matrix of the hypersurface;

[0017] Generate a hypersurface model based on the optimized coding matrix;

[0018] A data set model is constructed based on the continuity characterization parameter response of the metasurface model and the polarization angle and the quasi-polarization angle.

[0019] The specific method of completing the deep learning network training based on the data set model and using it to reversely design the polarization conversion metasurface to obtain the metasurface design solution is as follows:

[0020] Build a computer network model;

[0021] Construct a generative adversarial network architecture based on the generator network and the discriminator network;

[0022] Training the computer network model based on the data set model to obtain a training model;

[0023] Performing forward calculation on the metasurface coding distribution scheme using the training model to obtain a metasurface feature response;

[0024] Comparing the metasurface characteristic response with the demand response to obtain a response error;

[0025] By optimizing the generative adversarial network architecture, the hypersurface reverse design is performed and the hypersurface design scheme is obtained.

[0026] The specific method of performing reverse design of the hypersurface by optimizing the generative adversarial network architecture to obtain the hypersurface design scheme is as follows:

[0027] Inputting the demand response and random noise into the generator network to obtain the metasurface coding distribution scheme;

[0028] Inputting the hypersurface encoding distribution scheme and the data set model into the discriminator network to obtain the similarity of two types of model features;

[0029] Training the parameters of the generator network based on the response error and the feature similarity to obtain an optimized generator network;

[0030] Training the parameters of the discriminator network based on the feature similarity to obtain an optimized discriminator network;

[0031] The optimized generative adversarial network architecture constructed based on the optimized generator network and the optimized discriminator network is used to perform hypersurface reverse design to obtain diversified hypersurface reverse design solutions.

[0032] In a second aspect, the present invention provides a polarization conversion metasurface intelligent reverse design system based on deep learning, comprising a deep learning characterization module, a deep learning construction module, a deep learning training module and a polarization conversion metasurface reverse design module, wherein the deep learning characterization module, the deep learning construction module, the deep learning training module and the polarization conversion metasurface reverse design module are connected in sequence;

[0033] The deep learning characterization module is used to characterize the polarization conversion metasurface using polarization degree angle and quasi-polarization degree angle characterization mechanism to obtain a continuity characterization;

[0034] The deep learning building module builds a data set model, a computer network model and a generative adversarial network architecture based on the continuous representation;

[0035] The deep learning training module is used to train the computer network model and the generative adversarial network architecture to obtain a training model and an optimized generative adversarial network architecture respectively;

[0036] The polarization conversion metasurface reverse design module is used to use the training model to perform forward calculation on the metasurface coding to obtain the metasurface characteristic response, and to perform metasurface reverse design by optimizing the generative adversarial network architecture to obtain a metasurface design solution.

[0037] The present invention discloses an intelligent reverse design method for polarization conversion metasurface based on deep learning. The method characterizes the polarization conversion metasurface by using polarization degree angle and quasi-polarization degree angle characterization mechanism to obtain continuity characterization; constructs a data set model based on the continuity characterization; performs reverse design of the polarization conversion metasurface on the data set model to obtain a metasurface design scheme. The present invention realizes the functional characterization mechanism of the polarization conversion metasurface by using polarization degree angle and quasi-polarization degree angle to characterize the polarization conversion metasurface, thereby solving the problems of broadband phase difference discontinuity in traditional characterization methods and infinite characterization variables in axial ratio characterization methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 This is a flow chart of a method for intelligent reverse design of a polarization conversion metasurface based on deep learning provided by the present invention.

[0040] Figure 2 It is a schematic diagram of the structural arrangement of the coding metasurface unit.

[0041] Figure 3 It is a schematic diagram of the computer network cascade generative adversarial network architecture.

[0042] Figure 4 Schematic diagram of the computer network that maps the responses of the encoded metasurface.

[0043] Figure 5 This is a schematic diagram of the generative adversarial network architecture.

[0044] Figure 6 It is a structural schematic diagram of a deep learning-based polarization conversion metasurface intelligent reverse design system provided by the present invention.

[0045] 1-Deep learning characterization module, 2-Deep learning construction module, 3-Deep learning training module, 4-Polarization conversion metasurface reverse design module. DETAILED DESCRIPTION

[0046] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0047] See also Figures 1 to 5 In a first aspect, the present invention provides a method for intelligent inverse design of a polarization conversion metasurface based on deep learning, comprising the following steps:

[0048] S1 uses polarization angle and quasi-polarization angle characterization mechanism to characterize the polarization conversion metasurface and obtains continuity characterization;

[0049] Specifically, the polarization degree angle continuity is used to characterize the broadband linear-circular polarization conversion metasurface, and the quasi-polarization degree angle continuity is used to characterize the broadband cross-polarization conversion metasurface, thereby obtaining a continuity characterization.

[0050] The polarization angle has a value range of -π / 4 to π / 4, and when the polarization angle is ±π / 4, it indicates that the wave is left-handed circularly polarized;

[0051] The value range of the quasi-polarization angle is 0 to π / 2, and when the quasi-polarization angle is π / 2, it indicates that the wave is cross-polarized; when the quasi-polarization angle is 0, it indicates that the cross-polarization conversion function is not realized.

[0052] The relationship that can be expressed by the polarization angle β is:

[0053]

[0054] and

[0055]

[0056] Where e1 and e2 are the major and minor axes of the electric field polarization elliptical trajectory, respectively; where β ranges from -π / 4 to π / 4, and when β = ±π / 4, it indicates that the wave is left-handed circularly polarized. The polarization angle β can also characterize the linear-elliptical polarization conversion metasurface.

[0057] In the actual design, |β / π|≥0.2 is set as the criterion for circular polarization; the linear-elliptical polarization conversion is similarly based on the given |β / π| index requirement.

[0058] The relationship that can be expressed by the polarization angle β' is:

[0059]

[0060] Where |E ⊥ | and |E ∥ | respectively represent the amplitudes of the cross-polarization component and the co-polarization component; where β' ranges from 0 to π / 2, and when β'=π / 2, it means that the wave is cross-polarized; when β'=0, it means that the cross-polarization conversion function is not realized;

[0061] In actual design, β' / π≥0.4 is set as the criterion for cross-polarization.

[0062] S2 constructs a data set model based on the continuity representation;

[0063] The specific method is:

[0064] S21 uses electromagnetic simulation to obtain the metasurface propagation coefficient, and establishes the relationship between the polarization angle and the quasi-polarization angle parameter response and the metasurface propagation coefficient to obtain a metasurface continuity characterization relationship;

[0065] Specifically, CST full-wave electromagnetic software is used.

[0066] S22 describes the unit structure arrangement of the polarization conversion metasurface based on the coding matrix of the metasurface;

[0067] Specifically, Figure 2 The following are some schematic diagrams of second-order rotationally symmetric coding metasurfaces with a matrix size of 16×16. In applications, the metasurface with a coding matrix size of 16×16 discretizes the electromagnetic spatial parameters of the basic unit of the metasurface, that is, the coding degree of freedom of the metasurface is 2 256 , which can bring massive data models and bring huge freedom to the reverse design of super surface structures.

[0068] S23 iterates the arrangement of the unit structure using an iterative optimization algorithm based on the continuity representation relationship of the hypersurface to obtain an optimized coding matrix of the hypersurface;

[0069] S24 generates a hypersurface model based on the optimized coding matrix;

[0070] S25 constructs a data set model based on the continuity characterization parameter response of the metasurface model and the polarization angle and the quasi-polarization angle.

[0071] S3 completes deep learning network training based on the data set model, and uses it to perform reverse design of the polarization conversion metasurface to obtain a metasurface design solution.

[0072] The specific method is:

[0073] S31 builds a computer network model;

[0074] Specifically, the computer network model includes a multi-layer residual convolutional mapping network of a cascaded fully connected layer network (FCN), including an input layer, an output layer and a hidden layer, and adopts a Tanh() activation function.

[0075] S32 builds a generative adversarial network architecture based on the generator network and the discriminator network;

[0076] S33 trains the computer network model based on the data set model to obtain a training model;

[0077] Specifically, the input is the coded data set, and the output is the corresponding coded hypersurface discretization response. The coded hypersurface is quantized into a matrix as the input parameter of the computer network model; the coded hypersurface discretization response is quantized into a vector as the output parameter of the computer network model. The computer network model is trained according to the input parameters and output parameters, and the network parameters of the training model are finally fixed. The mapping relationship between the hypersurface and the corresponding response is obtained, and then the training model is obtained, which greatly shortens the calculation time.

[0078] S34 uses the training model to perform forward calculation on the metasurface coding distribution scheme to obtain a metasurface feature response;

[0079] S35 compares the metasurface characteristic response with the demand response to obtain a response error L R ;

[0080] Specifically, the response error L between the discretized output response of the computer network model and the demand response of the input generative adversarial network is R Backward guidance to update the parameters of generative adversarial network models.

[0081] S37 performs reverse design of the hypersurface by optimizing the generative adversarial network architecture and obtains the hypersurface design solution.

[0082] The specific method is:

[0083] S371 inputs the demand response and random noise into the generator network to obtain the metasurface coding distribution scheme;

[0084] S372 Input the hypersurface encoding distribution scheme and the data set model into the discriminator network to obtain the feature similarity L of the two types of models. C ;

[0085] S373 is based on the response error L R and the feature similarity L C Training parameters of the generator network to obtain an optimized generator network;

[0086] S374 Based on the feature similarity L C Training the parameters of the discriminator network to obtain an optimized discriminator network;

[0087] S375 optimizes the generative adversarial network architecture to perform reverse design of the hypersurface based on the optimized generator network and the optimized discriminator network to obtain diversified reverse design solutions of the hypersurface.

[0088] Specifically, by constructing a metasurface intelligent inverse design with an optimized generative adversarial network as the core, a variety of metasurface structures are obtained, which greatly reduces the complexity of the broadband polarization conversion metasurface design and improves the design efficiency, providing an efficient and intelligent technical approach for the design of diverse broadband polarization conversion metasurface solutions.

[0089] The present invention provides a deep learning-based intelligent reverse design method for polarization conversion metasurfaces. Firstly, a broadband polarization conversion metasurface functional characterization mechanism is realized by using polarization degree angle and quasi-polarization degree angle to characterize the broadband polarization conversion metasurface, thereby solving the phase difference discontinuity problem encountered in traditional characterization methods under broadband conditions and the problem of infinite characterization variables in the axial ratio characterization method. Secondly, a broadband polarization conversion data set model is established through an iterative optimization algorithm, thereby greatly shortening the calculation time. Finally, a diversified metasurface structure is obtained by constructing a metasurface intelligent reverse design with a generative adversarial network as the core, thereby realizing a user-defined, on-demand broadband polarization conversion metasurface design, thereby providing an efficient and intelligent technical approach for the design of diverse solutions for broadband polarization conversion metasurfaces.

[0090] The present invention not only solves the problem that the polarization state cannot be represented during the reverse design process, but also overcomes the problems of cumbersome calculations and limited effects brought about by the traditional forward design method of polarization conversion metasurfaces, providing an intelligent technical reference for the reverse design of polarization conversion metasurfaces.

[0091] See also Figure 6 In a second aspect, the present invention provides a polarization conversion metasurface intelligent reverse design system based on deep learning, comprising a deep learning characterization module 1, a deep learning construction module 2, a deep learning training module 3 and a polarization conversion metasurface reverse design module 4, wherein the deep learning characterization module 1, the deep learning construction module 2, the deep learning training module 3 and the polarization conversion metasurface reverse design module 4 are connected in sequence;

[0092] The deep learning characterization module 1 is used to characterize the polarization conversion metasurface using polarization angle and quasi-polarization angle characterization mechanism to obtain a continuity characterization;

[0093] The deep learning construction module 2 constructs a data set model, a computer network model and a generative adversarial network architecture based on the continuous representation;

[0094] The deep learning training module 3 is used to train the computer network model and the generative adversarial network architecture to obtain a training model and an optimized generative adversarial network architecture respectively;

[0095] The polarization conversion metasurface reverse design module 4 is used to use the training model to perform forward calculation on the metasurface coding to obtain the metasurface characteristic response, and perform metasurface reverse design by optimizing the generative adversarial network architecture to obtain a metasurface design solution.

[0096] What is disclosed above is only a preferred embodiment of the intelligent inverse design method of polarization conversion metasurface based on deep learning of the present invention. Of course, it cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiments are implemented, and the equivalent changes made according to the claims of the present invention are still within the scope of the invention.

Claims

1. A method for intelligent reverse design of polarization conversion metasurface based on deep learning, characterized in that: The following steps are involved: The polarization conversion metasurface is characterized by using polarization angle and quasi-polarization angle characterization mechanism to obtain a continuous characterization. constructing a data set model based on the continuity representation; Based on the data set model, deep learning network training is completed, and reverse design of polarization conversion metasurface is performed to obtain a metasurface design scheme; The specific method of constructing the data set model based on the continuity representation is: Electromagnetic simulation is used to obtain the metasurface propagation coefficient, and the relationship between the polarization angle and the quasi-polarization angle parameter response and the metasurface propagation coefficient is established to obtain a metasurface continuity characterization relationship; Describing the unit structure arrangement of the polarization conversion metasurface based on the encoding matrix of the metasurface; Iterating the arrangement of the unit structure using an iterative optimization algorithm based on the continuity representation relationship of the hypersurface to obtain an optimized coding matrix of the hypersurface; Generate a hypersurface model based on the optimized coding matrix; Constructing a data set model based on the metasurface model and the continuity characterization parameter response of the polarization angle and the quasi-polarization angle; The specific method of completing the deep learning network training based on the data set model and using it to reversely design the polarization conversion metasurface to obtain the metasurface design solution is: Build a computer network model; Construct a generative adversarial network architecture based on the generator network and the discriminator network; Training the computer network model based on the data set model to obtain a training model; Performing forward calculation on the metasurface coding distribution scheme using the training model to obtain a metasurface feature response; Comparing the metasurface characteristic response with the demand response to obtain a response error; By optimizing the generative adversarial network architecture, the hypersurface reverse design is performed and the hypersurface design scheme is obtained.

2. The method for intelligent reverse design of polarization conversion metasurface based on deep learning as claimed in claim 1, characterized in that: The specific method of using the polarization angle and quasi-polarization angle characterization mechanism to characterize the polarization conversion metasurface and obtain continuity characterization is: The polarization degree angle continuity is used to characterize the linear-circular polarization conversion metasurface, and the quasi-polarization degree angle continuity is used to characterize the cross-polarization conversion metasurface, thereby achieving the continuity characterization of the polarization conversion metasurface.

3. The method for intelligent reverse design of polarization conversion metasurface based on deep learning as claimed in claim 2, characterized in that: The polarization angle has a value range of -π / 4 to π / 4, and when the polarization angle is ±π / 4, it indicates that the wave is left-handed circularly polarized; The value range of the quasi-polarization angle is 0 to π / 2, and when the quasi-polarization angle is π / 2, it indicates that the wave is cross-polarized; when the quasi-polarization angle is 0, it indicates that the cross-polarization conversion function is not realized.

4. The method for intelligent reverse design of polarization conversion metasurface based on deep learning as claimed in claim 3, characterized in that: The specific method of performing reverse design of the hypersurface by optimizing the generative adversarial network architecture to obtain the hypersurface design scheme is as follows: Inputting the demand response and random noise into the generator network to obtain the metasurface coding distribution scheme; Inputting the hypersurface encoding distribution scheme and the data set model into the discriminator network to obtain the similarity of two types of model features; Training the parameters of the generator network based on the response error and the feature similarity to obtain an optimized generator network; Training the parameters of the discriminator network based on the feature similarity to obtain an optimized discriminator network; The optimized generative adversarial network architecture constructed based on the optimized generator network and the optimized discriminator network is used to perform hypersurface reverse design to obtain diversified hypersurface reverse design solutions.

5. A deep learning-based polarization conversion metasurface intelligent reverse design system, applied to the deep learning-based polarization conversion metasurface intelligent reverse design method according to claim 4, characterized in that: It includes a deep learning characterization module, a deep learning construction module, a deep learning training module and a polarization conversion metasurface reverse design module, wherein the deep learning characterization module, the deep learning construction module, the deep learning training module and the polarization conversion metasurface reverse design module are connected in sequence; The deep learning characterization module is used to characterize the polarization conversion metasurface using polarization degree angle and quasi-polarization degree angle characterization mechanism to obtain a continuity characterization; The deep learning building module builds a data set model, a computer network model and a generative adversarial network architecture based on the continuous representation; The deep learning training module is used to train the computer network model and the generative adversarial network architecture to obtain a training model and an optimized generative adversarial network architecture respectively; The polarization conversion metasurface reverse design module is used to use the training model to perform forward calculation on the metasurface coding to obtain the metasurface characteristic response, and to perform metasurface reverse design by optimizing the generative adversarial network architecture to obtain a metasurface design solution.

Images