Neural network optimization method, device and medium for realizing high uniformity optical waveguide

Abstract

The present invention relates to the field of neural network optimization technology, and provides a neural network optimization method, device and medium for realizing high uniformity optical waveguide, the method comprising: obtaining multiple super surface structures that meet the first design goal, and training and optimizing the preliminary generation adversarial network as a training set to obtain a target generation adversarial network; outputting a target device according to the target generation adversarial network; wherein the process of network training optimization comprises: setting preliminary incident light of three wavelengths of RGB, and then setting three different polar angles respectively, to obtain 9 incident lights; respectively determining the extended conductive light (1, 1) R, the outgoing light (0, 2) R and the outgoing light (-1, 1) R of each incident light, to obtain 27 optimization parameters; optimizing and iterating the preliminary generation adversarial network according to the 27 optimization parameters and multiple super surface structures. The present invention can quickly design a high uniformity optical waveguide through neural network optimization, reduce design time, and save computing resources.

Description

Technical Field

[0001] The present invention relates to the field of neural network optimization technology, and in particular to a neural network optimization method, device and medium for realizing high-uniformity optical waveguide. Background Art

[0002] Metasurface is a nanoscale optical component that can precisely control the phase, amplitude and polarization of light at a subwavelength scale. Compared with traditional optical devices, metasurfaces have thinner thickness and higher design freedom.

[0003] A key challenge in metasurface design is to develop algorithms that can effectively and efficiently produce high-performance devices. Traditional iterative optimization-based design methods can break through the performance limits of metasurfaces, but they require a lot of computing resources and take a long time, limiting their application to a small number of microscale devices. Summary of the invention

[0004] The present invention provides a neural network optimization method, device and medium for realizing high uniformity optical waveguide, so as to solve the problem that super surface design in the prior art takes a long time and consumes a lot of resources.

[0005] The present invention provides a neural network optimization method for realizing a high-uniformity optical waveguide, comprising: obtaining a plurality of supersurface structures that meet a first design goal; the emission uniformity value of the first design goal is a first preset value; using the plurality of supersurface structures as training sets to train and optimize a preliminary generative adversarial network to obtain a target generative adversarial network; outputting a first preset number of target devices according to the target generative adversarial network; the target devices are used as coupling-in gratings or coupling-out gratings; using the plurality of supersurface structures as training sets to train and optimize the preliminary generative adversarial network, comprising: setting preliminary incident light of three wavelengths of RGB; for each wavelength of preliminary incident light, respectively setting three different polar angles to obtain 9 incident lights; for each incident light, respectively determining extended conductive light (1, 1)R, emission light (0, 2)R and emission light (-1, 1)R to obtain 27 optimization parameters; optimizing and iterating the preliminary generative adversarial network according to the 27 optimization parameters and the plurality of supersurface structures, the objective function of the preliminary generative adversarial network during the optimization iteration is:

[0006] (1,1)R(θ,φ)×(-1,1)R(θ2,φ2)=(0,2)R(θ,φ);

[0007] Wherein, (1,1)R is the extended transmission light generated by the first incident on the grating, (0,2)R is the outgoing light generated by the first incident on the grating; (-1,1)R is the outgoing light generated by the diffraction of the extended transmission light (1,1)R after the first incident on the grating; θ represents the polar angle of the first incident light; φ represents the azimuth angle of the first incident light;

[0008] θ2 represents the polar angle of the second incident light; φ2 represents the azimuth angle of the second incident light.

[0009] According to a neural network optimization method for realizing high uniformity optical waveguide provided by the present invention, a plurality of metasurface structures are used as training sets to train and optimize a preliminary generative adversarial network to obtain a target generative adversarial network, and the method further includes: a generator of the preliminary generative adversarial network generates a plurality of preliminary structures; determining the extended conductive light (1, 1) R, the emitted light (0, 2) R and the emitted light (-1, 1) R of different incident lights in the plurality of preliminary structures; and determining the extended conductive light (1, 1) R, the emitted light (0, 2) R and the emitted light (-1, 1) R of different incident lights in the plurality of preliminary structures. The diffraction efficiency of the corresponding preliminary structure is (1, 1)R×(-1, 1)R / (0, 2)R according to the output light (-1, 1)R; the preliminary generative adversarial network is optimized and iterated according to the diffraction efficiency (1, 1)R×(-1, 1)R / (0, 2)R, and when the discriminator of the preliminary generative adversarial network determines that the diffraction efficiencies (1, 1)R×(-1, 1)R / (0, 2)R of the second preset number of preliminary structures all satisfy the objective function, it is determined that the training is completed and the target generative adversarial network is obtained.

[0010] According to a neural network optimization method for realizing high uniformity optical waveguide provided by the present invention, a first preset number of target devices is outputted according to a target generative adversarial network, including: outputting a first number of intermediate devices according to the target generative adversarial network; determining a first preset number of target devices that meet a second design target from the first number of intermediate devices; wherein the first number is greater than the first preset number; and the first design target and the second design target are different.

[0011] According to a neural network optimization method for realizing high-uniformity optical waveguide provided by the present invention, a plurality of metasurface structures meeting a first design goal are obtained, including: using an electromagnetic solver and an optimization algorithm to obtain a second number of metasurface structures meeting the first design goal; wherein the second number is less than the first preset number.

[0012] According to a neural network optimization method for realizing high uniformity optical waveguide provided by the present invention, the optimization algorithm is a companion topology optimization algorithm or a particle swarm optimization algorithm.

[0013] According to a neural network optimization method for realizing high uniformity optical waveguide provided by the present invention, a first preset number of target devices are output according to a target generative adversarial network, including: inputting fixed parameters into the target generative adversarial network to obtain the first preset number of target devices output by the target generative adversarial network; wherein the fixed parameters refer to the same parameters in all target devices; the fixed parameters include at least one of the type of incident light, the period of the outcoupling grating in the x-direction, the period of the outcoupling grating in the y-direction, and the height of the outcoupling grating.

[0014] The present invention also provides a neural network optimization device for realizing high-uniformity optical waveguide, comprising: a metasurface structure module, used to obtain multiple metasurface structures that meet a first design goal; the emission uniformity value of the first design goal is a first preset value; a network training optimization module, used to train and optimize a preliminary generative adversarial network using multiple metasurface structures as training sets to obtain a target generative adversarial network; a target device module, used to output a first preset number of target devices according to the target generative adversarial network; the target device is used as an in-coupled grating or an out-coupled grating; the network training optimization module is used to set preliminary incident light of three wavelengths of RGB; for each wavelength of preliminary incident light, three different polar angles are set respectively to obtain 9 incident lights; for each incident light, the extended conductive light (1, 1) R, the emitted light (0, 2) R and the emitted light (-1, 1) R are determined respectively to obtain 27 optimization parameters; according to the 27 optimization parameters and multiple metasurface structures, the preliminary generative adversarial network is optimized and iterated, and the objective function of the preliminary generative adversarial network during the optimization iteration is:

[0015] (1,1)R(θ,φ)×(-1,1)R(θ2,φ2)=(0,2)R(θ,φ);

[0016] Wherein, (1,1)R is the extended transmission light generated by the first incident on the grating, (0,2)R is the outgoing light generated by the first incident on the grating; (-1,1)R is the outgoing light generated by the diffraction of the extended transmission light (1,1)R after the first incident on the grating; θ represents the polar angle of the first incident light; φ represents the azimuth angle of the first incident light;

[0017] θ2 represents the polar angle of the second incident light; φ2 represents the azimuth angle of the second incident light.

[0018] The present invention also provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, any one of the above-mentioned neural network optimization methods for realizing a high-uniformity optical waveguide is implemented.

[0019] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the computer program implements any of the above-mentioned neural network optimization methods for realizing a high-uniformity optical waveguide.

[0020] The present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-mentioned neural network optimization methods for realizing a high-uniformity optical waveguide.

[0021] The neural network optimization method, device and medium for realizing high uniformity optical waveguide provided by the present invention include: obtaining multiple super surface structures that meet the first design target; the emission uniformity value of the first design target is a first preset value; using multiple super surface structures as training sets to train and optimize the preliminary generation adversarial network to obtain a target generation adversarial network; outputting a first preset number of target devices according to the target generation adversarial network; the target device is used as a coupling grating or an coupling grating; using multiple super surface structures as training sets to train and optimize the preliminary generation adversarial network, including: setting preliminary incident light of three wavelengths of RGB; for each wavelength of preliminary incident light, three different polar angles are set respectively to obtain 9 incident lights; for each incident light, the extended transmission light (1, 1) R, the output light (0, 2) R and the output light (-1, 1) R are determined respectively to obtain 27 optimization parameters; according to the 27 optimization parameters and multiple super surface structures, the preliminary generation adversarial network is optimized and iterated. Through the above-mentioned method, the present invention can quickly design a wide-band, large-angle, high-uniformity optical waveguide through neural network optimization, reduce design time, and save computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are 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.

[0023] Figure 1 It is a flow chart of a neural network optimization method for realizing a high uniformity optical waveguide provided in an embodiment of the present invention.

[0024] Figure 2 It is a schematic diagram of the process of preliminary generative adversarial network training optimization provided by an embodiment of the present invention.

[0025] Figure 3 It is a schematic diagram of the first diffraction light path provided by an embodiment of the present invention.

[0026] Figure 4 It is a schematic diagram of the second diffraction light path provided by an embodiment of the present invention.

[0027] Figure 5 It is a schematic diagram of the training process of the generator and the discriminator in the generative adversarial network provided by an embodiment of the present invention.

[0028] Fig. 6A This is one of the schematic diagrams of the binary image of the out-coupled grating output by the target generative adversarial network provided in an embodiment of the present invention.

[0029] Figure 6BThe embodiment of the present invention provides Fig. 6A Schematic diagram of the corresponding loss function curve changing with the number of iterations.

[0030] Fig. 7A This is the second schematic diagram of the binary image of the out-coupled grating output by the target generative adversarial network provided in an embodiment of the present invention.

[0031] Figure 7B The embodiment of the present invention provides Fig. 7A Schematic diagram of the corresponding loss function changing with the number of iterations.

[0032] Figure 8 It is a structural schematic diagram of a neural network optimization device for realizing a high uniformity optical waveguide provided in an embodiment of the present invention.

[0033] Fig. 9 It is a schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0035] In the description of the embodiments of the present invention, it should be noted that the terms "first", "second", and "third" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance. The description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the embodiments of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples without contradiction.

[0036] The present invention provides a neural network optimization method for realizing high uniformity optical waveguide, please refer to Figure 1-Figure 2 , Figure 1 is a flow chart of a neural network optimization method for realizing a high uniformity optical waveguide provided by an embodiment of the present invention, Figure 21 is a schematic diagram of a preliminary generative adversarial network training optimization process provided by an embodiment of the present invention. In this embodiment, a neural network optimization method for realizing a high uniformity optical waveguide may include steps S110 to S130, each of which is specifically as follows:

[0037] S110: Acquire multiple metasurface structures that meet the first design goal; the emission uniformity value of the first design goal is a first preset value.

[0038] S120: Using the multiple hypersurface structures as training sets to train and optimize the preliminary generative adversarial network to obtain a target generative adversarial network.

[0039] S130: Outputting a first preset number of target devices according to the target generative adversarial network; the target devices are used as coupling-in gratings or coupling-out gratings.

[0040] Specifically, the step of using multiple hypersurface structures as training sets to train and optimize the preliminary generative adversarial network may include the following steps:

[0041] S121: Setting preliminary incident light of three wavelengths of RGB.

[0042] S122: For each wavelength of the initial incident light, three different polar angles are set respectively to obtain 9 incident lights.

[0043] S123: For each incident light, determine the extended guided light (1, 1)R, the outgoing light (0, 2)R and the outgoing light (-1, 1)R respectively, and obtain 27 optimization parameters.

[0044] S124: Optimize and iterate the preliminary generative adversarial network according to 27 optimization parameters and multiple hypersurface structures.

[0045] In this embodiment, the objective function of the preliminary generative adversarial network during optimization iteration is:

[0046] (1,1)R(θ,φ)×(-1,1)R(θ2,φ2)=(0,2)R(θ,φ).

[0047] Among them, (1, 1)R is the extended transmission light generated by the first incident on the grating, (0, 2)R is the output light generated by the first incident on the grating; (-1, 1)R is the extended transmission light generated after the first incident is diffracted, and the output light is diffracted after (1, 1)R is incident on the grating again; θ represents the polar angle of the first incident light; φ represents the azimuth angle of the first incident light; θ2 represents the polar angle of the second incident light; φ2 represents the azimuth angle of the second incident light.

[0048] In this embodiment, in order to obtain a highly uniform target metasurface device as an in-coupled grating or an out-coupled grating, a metasurface structure that satisfies the emission uniformity is used as a training set. Specifically, a plurality of metasurface structures meet the first design goal, and the emission uniformity value of the first design goal is a first preset value.

[0049] Optionally, the first preset value is in the range of 0.9-1.1. Preferably, when the first preset value is 1, the emission uniformity of the metasurface structure is the best.

[0050] Conditionally constrained generative adversarial neural networks (GANs) can provide a good design solution in optical design, which can be used to generate regular or irregular shaped photonic device structures (where regular shapes refer to collective shapes such as cylinders, cubes, and ellipses).

[0051] In this embodiment, a plurality of metasurface structures that meet the first design goal are used to form a training set, a generative adversarial network is trained, and the parameters of the generative adversarial network are adjusted to obtain a target generative adversarial network and save it. The saved target generative adversarial network can be used to generate any number of device structures that meet the first design goal, which can reduce design time and save computing resources compared to traditional design methods.

[0052] For example, the initial incident light of the three wavelengths of RGB are red light wavelength (e.g., 620nm), green light wavelength (e.g., 530nm) and blue light wavelength (e.g., 460nm). Those skilled in the art can specifically design three different polar angles according to actual conditions, for example, they can be set to 50°, 60° and 70°.

[0053] In this embodiment, the emission uniformity of the target device is considered based on the first diffraction and the second diffraction.

[0054] See also Figure 3 and Figure 4 , Figure 3 is a schematic diagram of the first diffraction light path provided by an embodiment of the present invention; Figure 4 Schematic diagram of the second diffraction light path provided by an embodiment of the present invention. The light path schematic diagrams are all displayed based on the xyz coordinate axis. represents the grazing incidence angle, Represents the deflection angle of diffraction on both sides.

[0055] Furthermore, the step of outputting a first preset number of target devices according to the target generative adversarial network specifically includes:

[0056] Output a first number of intermediate components according to the target generative adversarial network; determine a first preset number of target components that meet the second design target from the first number of intermediate components; wherein the first number is greater than the first preset number; and the first design target and the second design target are different.

[0057] In this embodiment, optical devices that meet multiple design goals can also be obtained. Specifically, according to the target generative adversarial network, a first number of intermediate devices are output, wherein the intermediate devices are optical devices that meet the first design goal. On this basis, target devices that meet the second design goal are further selected from the first number of intermediate devices. The first design goal and the second design goal are different. The second design goal can be requirements for the band, angle, manufacturing process, etc., so as to obtain a wide-band, large-angle, high-uniformity and easy-to-manufacture out-coupling grating or in-coupling grating.

[0058] Based on any of the above embodiments, the step of obtaining a plurality of metasurface structures that meet the first design goal specifically includes:

[0059] An electromagnetic solver and an optimization algorithm are used to obtain a second number of metasurface structures that meet the first design goal; wherein the second number is less than the first preset number.

[0060] Optionally, the optimization algorithm is a companion topology optimization algorithm or a particle swarm optimization algorithm.

[0061] In this embodiment, an electromagnetic solver (such as FDTD, RCWA, COMSOL) can be used in combination with a certain algorithm (adjoint topology optimization, particle swarm optimization, etc.) to design a certain number of device structures that meet the design objectives. COMSOL is a multi-physics simulation software platform.

[0062] The adjoint topology optimization algorithm is used in conjunction with the RCWA (Rigorous Coupled-Wave Analysis) solver. RCWA is an electromagnetic simulation algorithm with high accuracy and fast calculation speed. The RCWA algorithm can be used to calculate and establish a database. Compared with FEM (Finite Element Method) and FDTD (Finite-Difference Time-Domain), RCWA is fast and accurate.

[0063] RCWA calculates S parameters to obtain the complex amplitudes of the transmitted electric field and the reflected electric field. It can obtain the transmitted amplitude and phase distribution, and can also calculate the electromagnetic field in the structure.

[0064] For the distribution of complex arbitrary-shaped unit structures, if you use software such as FDTD to get more accurate results, it will consume a lot of computing resources in simulation time and meshing. However, using the semi-analytical RCWA algorithm, you can get results quickly. For the simulation calculation of irregular-shaped metasurfaces, since the unit structure has a finer resolution, it is necessary to adjust the mesh accuracy of the electromagnetic solver to a relatively high level. Using FDTD will consume huge computing resources and time. Therefore, the electromagnetic solvers accompanying topology optimization now all use the RCWA algorithm.

[0065] Classical adjoint topology optimization reverse design requires calculation of the gradient of dielectric constant, so the adjoint method is used, which can be regarded as two electromagnetic simulation calculations, but it takes a long time to optimize the structure. If it is iterated 300 times, it takes about 20 hours. Therefore, it takes a lot of time to prepare a sufficient number of training sets. With the development of machine learning, electromagnetic simulation algorithms can record the gradient information of each step in the simulation process, which can avoid the consumption of computing resources of the adjoint algorithm.

[0066] Automatic Differentiation Adjoint Topology Optimization (AD-ATO) is a topology optimization method that combines automatic differentiation technology and adjoint methods. The combination of automatic differentiation and adjoint methods significantly reduces the overhead of calculating gradients and improves optimization efficiency. However, these optimization algorithms require a large amount of computing resources to achieve the output of the target device. Therefore, in this embodiment, only a small number of multiple hypersurface structures that meet the design goals are generated by the optimization algorithm as a training set to train and optimize the adversarial network, thereby obtaining a target generative adversarial network, and then using the target generative adversarial network to output the first preset number of target devices.

[0067] Based on any of the above embodiments, the steps of using multiple hypersurface structures as training sets to train and optimize the preliminary generative adversarial network to obtain a target generative adversarial network specifically include:

[0068] The generator of the preliminary generative adversarial network generates multiple preliminary structures; determines the extended transmission light (1, 1) R, the output light (0, 2) R and the output light (-1, 1) R of different incident lights in the multiple preliminary structures; determines the diffraction efficiency (1, 1) R×(-1, 1) R / (0, 2) R of the corresponding preliminary structures based on the extended transmission light (1, 1) R, the output light (0, 2) R and the output light (-1, 1) R of different incident lights in the multiple preliminary structures; optimizes and iterates the preliminary generative adversarial network according to the diffraction efficiency (1, 1) R×(-1, 1) R / (0, 2) R, and when the discriminator of the preliminary generative adversarial network determines that the diffraction efficiencies (1, 1) R×(-1, 1) R / (0, 2) R of the second preset number of preliminary structures all satisfy the objective function, it is determined that the training is completed, and a target generative adversarial network is obtained.

[0069] Generative Adversarial Networks (GAN) is a deep learning model that mainly consists of two neural networks: the generator and the discriminator, which are trained through an adversarial process.

[0070] The main task of the generator is to generate data samples that are as realistic as possible. The main task of the discriminator is to distinguish whether the input data comes from the real data set or the fake data generated by the generator. It receives the real data and the fake data generated by the generator, and judges through the neural network layer, and outputs a probability value, which indicates the probability that the input data is real. The generator and the discriminator compete with each other during the training process. The generator tries to generate more and more realistic fake data to deceive the discriminator, while the discriminator tries to continuously improve its ability to distinguish between real and fake data.

[0071] In some embodiments, more than 300 metasurface structures can be obtained through the topology optimization algorithm. By analyzing the actual diffraction efficiency and pattern comparison, 39 metasurface structures with better effects, 39 metasurface structures and 161 metasurface structures with average effects are selected to make the first training set, with a total of 200 data, divided into two labels. Note that in order to facilitate GAN data processing, the original hypersurface model has a pixel size of 530×918 and has a high symmetry. Only 1 / 4 of the original image is taken, and the pixels are modified to 32×128.

[0072] See also Figure 5 , Figure 5 It is a schematic diagram of the training process of the generator and the discriminator in the generative adversarial network provided by an embodiment of the present invention.

[0073] Based on any of the above embodiments, the step of outputting a first preset number of target devices according to the target generative adversarial network specifically includes:

[0074] Fixed parameters are input into a target generative adversarial network to obtain a first preset number of target devices output by the target generative adversarial network; wherein the fixed parameters refer to the same parameters in all target devices; the fixed parameters include at least one of the type of incident light, the period of the outcoupling grating in the x-direction, the period of the outcoupling grating in the y-direction, and the height of the outcoupling grating.

[0075] In this embodiment, fixed parameters are set as an input of the generator. The generator generates the design or characteristics of the target device according to these fixed parameters and other variable parameters, and finally obtains a batch of target devices with specific parameters.

[0076] For example, the period of the out-coupled grating in the target generative adversarial network is 918nm in the x-direction, the period of the out-coupled grating in the y-direction is 530nm, and the height is 80nm.

[0077] At the three wavelengths of RGB, the incident light I(θ, φ) (polar angle θ = 50°, 60°, 70°, azimuth angle φ = 0°); (1, 1) R(θ, φ) × (-1, 1) R(θ 2， φ2)=(0,2)R(θ,φ), that is, the diffraction light efficiency of (1,1)R multiplied by the diffraction light efficiency of (-1,1)R is equal to the diffraction light efficiency of (0,2)R.

[0078] (1,1)R and (0,2)R are the diffracted lights generated by the first incident on the grating, with incident angles of 50°, 60° and 70° respectively, where (1,1)R is the extended guided light and (0,2)R is the outgoing light.

[0079] (-1, 1)R is the extended guided light generated after the first incident light is diffracted, and (1, 1)R is the outgoing light generated after it is incident on the grating again and diffracted.

[0080] See also Fig. 6A and Figure 6B , Fig. 6A It is one of the binary image schematic diagrams of the out-coupled grating output by the target generative adversarial network provided in an embodiment of the present invention, wherein the horizontal and vertical coordinates are lengths in nm, gray is 1 representing the structure, and black is 0 representing the substrate. Figure 6B The embodiment of the present invention provides Fig. 6A The corresponding loss function changes with the number of iterations. The following table shows the relevant optimization parameters and uniformity values:

[0081]

[0082] See also Fig. 7A and Figure 7B , Fig. 7AThis is the second schematic diagram of the binary image of the out-coupled grating output by the target generative adversarial network provided in an embodiment of the present invention, wherein the horizontal and vertical axes are lengths in nm, gray is 1 representing the structure, and black is 0 representing the substrate. Figure 7B The embodiment of the present invention provides Fig. 7A The corresponding loss function changes with the number of iterations. The following table shows the relevant optimization parameters and uniformity values:

[0083]

[0084] The present invention also provides a neural network optimization device for realizing a high uniformity optical waveguide. The neural network optimization device for realizing a high uniformity optical waveguide provided by the present invention is described below. The neural network optimization device for realizing a high uniformity optical waveguide described below and the neural network optimization method for realizing a high uniformity optical waveguide described above can be referenced to each other.

[0085] See also Figure 8 , Figure 8 is a schematic diagram of the structure of a neural network optimization device for realizing a high uniformity optical waveguide provided by an embodiment of the present invention. In this embodiment, the neural network optimization device for realizing a high uniformity optical waveguide may include:

[0086] The super surface structure module 810 is used to obtain multiple super surface structures that meet the first design goal; the emission uniformity value of the first design goal is a first preset value.

[0087] The network training optimization module 820 is used to train and optimize the preliminary generative adversarial network using multiple hypersurface structures as training sets to obtain a target generative adversarial network.

[0088] The target device module 830 is used to output a first preset number of target devices according to the target generative adversarial network; the target devices are used as coupling-in gratings or coupling-out gratings.

[0089] Among them, the network training optimization module 820 is also used to set the preliminary incident light of three wavelengths, RGB; for the preliminary incident light of each wavelength, three different polar angles are set respectively to obtain 9 incident lights; for each incident light, the extended conductive light (1, 1)R, the outgoing light (0, 2)R and the outgoing light (-1, 1)R are determined respectively to obtain 27 optimization parameters; according to the 27 optimization parameters and multiple hypersurface structures, the preliminary generative adversarial network is optimized and iterated.

[0090] The objective function of the initial generation adversarial network during optimization iteration is:

[0091] (1,1)R(θ,φ)×(-1,1)R(θ2,φ2)=(0,2)R(θ,φ);

[0092] Wherein, (1,1)R is the extended transmission light generated by the first incident on the grating, (0,2)R is the outgoing light generated by the first incident on the grating; (-1,1)R is the outgoing light generated by the diffraction of the extended transmission light (1,1)R after the first incident on the grating; θ represents the polar angle of the first incident light; φ represents the azimuth angle of the first incident light; represents the polar angle of the second incident light; φ2 represents the azimuth angle of the second incident light.

[0093] Based on any of the above embodiments, the network training optimization module 820 can also be specifically used for: the generator of the preliminary generative adversarial network generates multiple preliminary structures; determines the extended transmission light (1, 1) R, the output light (0, 2) R and the output light (-1, 1) R of different incident lights in the multiple preliminary structures; determines the diffraction efficiency (1, 1) R×(-1, 1) R / (0, 2) R of the corresponding preliminary structure based on the extended transmission light (1, 1) R, the output light (0, 2) R and the output light (-1, 1) R of different incident lights in the multiple preliminary structures; optimizes and iterates the preliminary generative adversarial network according to the diffraction efficiency (1, 1) R×(-1, 1) R / (0, 2) R, and when the discriminator of the preliminary generative adversarial network determines that the diffraction efficiencies (1, 1) R×(-1, 1) R / (0, 2) R of the second preset number of preliminary structures all satisfy the objective function, it is determined that the training is completed and the target generative adversarial network is obtained.

[0094] Based on any of the above embodiments, the target device module 830 can be specifically used to: output a first number of intermediate devices according to a target generative adversarial network; determine a first preset number of target devices that meet a second design target from the first number of intermediate devices; wherein the first number is greater than the first preset number; and the first design target and the second design target are different.

[0095] Based on any of the above embodiments, the metasurface structure module 810 can be specifically used to: use an electromagnetic solver and an optimization algorithm to obtain a second number of metasurface structures that meet the first design goal; wherein the second number is less than the first preset number.

[0096] Based on any of the above embodiments, the optimization algorithm is a companion topology optimization algorithm or a particle swarm optimization algorithm.

[0097] Based on any of the above embodiments, the target device module 830 can be specifically used to: input fixed parameters into the target generative adversarial network to obtain a first preset number of target devices output by the target generative adversarial network; wherein the fixed parameters refer to the same parameters in all target devices; the fixed parameters include at least one of the type of incident light, the period of the out-coupling grating in the x-direction, the period of the out-coupling grating in the y-direction, and the height of the out-coupling grating.

[0098] On the other hand, an embodiment of the present invention further provides an electronic device, see Fig. 9 , Fig. 9 is a schematic diagram of the physical structure of an electronic device provided by an embodiment of the present invention, such as Fig. 9 As shown, the electronic device may include a memory 920, a processor 910, and a computer program stored in the memory 920 and executable on the processor 910. When the processor 910 executes the program, the neural network optimization method for realizing a high uniformity optical waveguide provided by the above methods is implemented.

[0099] Optionally, the electronic device may further include a communication bus 930 and a communication interface (CommunicationsInterface) 940, wherein the processor 910, the communication interface 940, and the memory 920 communicate with each other through the communication bus 930. The processor 910 may call a computer program in the memory 920 to execute a neural network optimization method for realizing a high uniformity optical waveguide, and the method may include:

[0100] A plurality of metasurface structures meeting the first design goal are obtained; the emission uniformity value of the first design goal is a first preset value; the plurality of metasurface structures are used as training sets to train and optimize the preliminary generative adversarial network to obtain a target generative adversarial network; a first preset number of target devices are output according to the target generative adversarial network; the target device is used as an in-coupled grating or an out-coupled grating; the plurality of metasurface structures are used as training sets to train and optimize the preliminary generative adversarial network, including: setting preliminary incident light of three wavelengths, RGB; for each wavelength of preliminary incident light, three different polar angles are set respectively to obtain 9 incident lights; for each incident light, the extended guided light (1, 1)R, the exit light (0, 2)R and the exit light (-1, 1)R are determined respectively to obtain 27 optimization parameters; the preliminary generative adversarial network is optimized and iterated according to the 27 optimization parameters and the plurality of metasurface structures, and the objective function of the preliminary generative adversarial network during the optimization iteration is:

[0101] (1,1)R(θ,φ)×(-1,1)R(θ2,φ2)=(0,2)R(θ,φ);

[0102] Wherein, (1,1)R is the extended transmission light generated by the first incident on the grating, (0,2)R is the outgoing light generated by the first incident on the grating; (-1,1)R is the outgoing light generated by the diffraction of the extended transmission light (1,1)R after the first incident on the grating; θ represents the polar angle of the first incident light; φ represents the azimuth angle of the first incident light; represents the polar angle of the second incident light; φ2 represents the azimuth angle of the second incident light.

[0103] In addition, the logic instructions in the above-mentioned memory 920 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0104] On the other hand, the present invention also provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the neural network optimization method for realizing high-uniformity optical waveguide provided by the above methods. The steps and principles have been introduced in detail in the above methods and will not be repeated here.

[0105] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it is implemented to execute the neural network optimization method for realizing high-uniformity optical waveguide provided by the above methods. The steps and principles thereof have been introduced in detail in the above methods and will not be repeated here.

[0106] Non-transitory computer-readable storage media can be any available media or data storage devices that can be accessed by the processor, including but not limited to magnetic storage (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MO)), optical storage (such as CDs, DVDs, BDs, HVDs, etc.), and semiconductor storage (such as ROM, EPROM, EEPROM, non-volatile memory (NANDFLASH), solid-state drives (SSDs)), etc.

[0107] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative labor.

[0108] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A neural network optimization method for realizing high uniformity optical waveguide, characterized in that: include: Acquire a plurality of metasurface structures that meet the first design goal; The emission uniformity value of the first design target is a first preset value; Using the multiple hypersurface structures as training sets to train and optimize the preliminary generative adversarial network, to obtain a target generative adversarial network; Outputting a first preset number of target devices according to the target generative adversarial network; The target device is used as an in-coupling grating or an out-coupling grating; The method further comprises: using the plurality of hypersurface structures as training sets to train and optimize the preliminary generative adversarial network, comprising: Set the initial incident light of three wavelengths of RGB; For each wavelength of initial incident light, three different polar angles are set to obtain 9 incident lights; For each incident light, the extended guided light (1, 1) R, the outgoing light (0, 2) R and the outgoing light (-1, 1) R are determined respectively, and 27 optimization parameters are obtained; The preliminary generative adversarial network is optimized and iterated according to the 27 optimization parameters and the multiple hypersurface structures, and the objective function of the preliminary generative adversarial network during the optimization iteration is: (1,1)R(θ,φ)×(-1,1)R(θ2,φ2)=(0,2)R(θ,φ); Among them, (1, 1)R is the extended transmission light generated by the first incident on the grating, (0, 2)R is the output light generated by the first incident on the grating; (-1, 1)R is the extended transmission light generated after the first incident is diffracted, and the output light is diffracted after (1, 1)R is incident on the grating again; θ represents the polar angle of the first incident light; φ represents the azimuth angle of the first incident light; θ2 represents the polar angle of the second incident light; φ2 represents the azimuth angle of the second incident light.

2. The neural network optimization method for realizing high uniformity optical waveguide according to claim 1, characterized in that: The method further comprises: using the plurality of hypersurface structures as training sets to train and optimize the preliminary generative adversarial network to obtain a target generative adversarial network; The generator of the preliminary generative adversarial network generates a plurality of preliminary structures; Determine the extended guided light (1, 1)R, the outgoing light (0, 2)R and the outgoing light (-1, 1)R of different incident lights in a plurality of preliminary structures; Determine the diffraction efficiency (1,1)R×(-1,1)R / (0,2)R of the corresponding preliminary structure based on the extended guided light (1,1)R, the outgoing light (0,2)R and the outgoing light (-1,1)R of different incident lights in the plurality of preliminary structures; The preliminary generative adversarial network is optimized iteratively according to the diffraction efficiency (1, 1)R×(-1, 1)R / (0, 2)R. When the discriminator of the preliminary generative adversarial network determines that the diffraction efficiencies (1, 1)R×(-1, 1)R / (0, 2)R of a second preset number of preliminary structures all satisfy the objective function, the training is determined to be completed, and the target generative adversarial network is obtained.

3. The neural network optimization method for realizing high uniformity optical waveguide according to claim 1, characterized in that: Outputting a first preset number of target devices according to the target generation adversarial network includes: Generate an adversarial network outputting a first number of intermediate devices according to the target; Determine a first preset number of target devices that meet a second design goal from the first number of intermediate devices; wherein the first number is greater than the first preset number; and the first design goal and the second design goal are different.

4. The neural network optimization method for realizing high uniformity optical waveguide according to claim 1, characterized in that: The obtaining of a plurality of super-surface structures that meet the first design goal comprises: An electromagnetic solver and an optimization algorithm are used to obtain a second number of metasurface structures that meet the first design goal; wherein the second number is less than the first preset number.

5. The neural network optimization method for realizing high uniformity optical waveguide according to claim 4, characterized in that: The optimization algorithm is a companion topology optimization algorithm or a particle swarm optimization algorithm.

6. The neural network optimization method for realizing high uniformity optical waveguide according to any one of claims 1 to 5, characterized in that: Outputting a first preset number of target devices according to the target generation adversarial network includes: Fixed parameters are input into the target generative adversarial network to obtain a first preset number of target devices output by the target generative adversarial network; wherein the fixed parameters refer to the same parameters in all target devices; the fixed parameters include at least one of the type of incident light, the period of the outcoupling grating in the x-direction, the period of the outcoupling grating in the y-direction and the height of the outcoupling grating.

7. A neural network optimization device for realizing high uniformity optical waveguide, characterized in that: include: A super surface structure module, used to obtain a plurality of super surface structures that meet the first design goal; The emission uniformity value of the first design target is a first preset value; A network training optimization module, used to train and optimize the preliminary generative adversarial network using the multiple hypersurface structures as training sets to obtain a target generative adversarial network; A target device module, configured to output a first preset number of target devices according to the target generative adversarial network; the target devices are used as coupling-in gratings or coupling-out gratings; The network training optimization module is used to set preliminary incident light of three wavelengths, RGB; for each wavelength of preliminary incident light, three different polar angles are set respectively to obtain 9 incident lights; for each incident light, the extended transmission light (1, 1) R, the outgoing light (0, 2) R and the outgoing light (-1, 1) R are determined respectively to obtain 27 optimization parameters; according to the 27 optimization parameters and the multiple super surface structures, the preliminary generative adversarial network is optimized and iterated, and the objective function of the preliminary generative adversarial network during the optimization iteration is: (1,1)R(θ,φ)×(-1,1)R(θ2,φ2)=(0,2)R(θ,φ); Among them, (1, 1)R is the extended transmission light generated by the first incident on the grating, and (0, 2)R is the outgoing light generated by the first incident on the grating; (-1, 1)R is the extended guided light produced after the first incident light is diffracted, and the outgoing light is diffracted after (1, 1)R is incident on the grating again; θ represents the polar angle of the first incident light; φ represents the azimuth angle of the first incident light; θ2 represents the polar angle of the second incident light; φ2 represents the azimuth angle of the second incident light.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that: When the processor executes the computer program, the neural network optimization method for realizing a high-uniformity optical waveguide as claimed in any one of claims 1 to 6 is implemented.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the neural network optimization method for realizing a high-uniformity optical waveguide as claimed in any one of claims 1 to 6 is implemented.

10. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the neural network optimization method for realizing a high-uniformity optical waveguide as claimed in any one of claims 1 to 6 is implemented.

Images