A non-local generative scatterer parameter design method
By establishing a mapping relationship between scatterer parameters and electric field, and using generative adversarial networks to optimize the polarizability parameters of the scatterer array, the problems of high complexity and nonlocal effects in existing metasurface designs are solved, and efficient design of arbitrary electric field distributions is realized.
Patent Information
- Application Number
- CN202210722395.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-17
- Publication Date
- 2025-10-17
- Estimated Expiration
- 2042-06-17
AI Technical Summary
Existing metasurface design methods cannot achieve arbitrary electric field distributions or consider nonlocal effects between metamaterial units, resulting in high design complexity and the inability to achieve global optimization.
A nonlocal generative scatterer parameter design method is adopted. By establishing the mapping relationship between scatterer parameters and electric field, the polarizability parameters of the scatterer array are optimized using a generative adversarial network architecture, thereby realizing the global design of arbitrary electric field distribution.
The design of arbitrary electric field distributions was realized, which reduced the design time of metasurfaces, improved design efficiency, and explained the nonlocal response between scatterers. The average peak signal-to-noise ratio between the generated scatterer parameters and the target electric field distribution reached 103.1464 dB.
Smart Images
Figure CN115203904B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of electronic information, and particularly relates to a non-local generative scatterer parameter design method. BACKGROUND
[0002] Metamaterial is a kind of artificial material with sub-wavelength structure arranged periodically, which has the ability to control the amplitude, phase and polarization of electromagnetic waves. Its typical representatives include frequency selective surfaces (FSS) and electromagnetic band gap structures (EBG). Metasurface, as a two-dimensional metamaterial, has the advantage of occupying less physical space. Its main applications include electromagnetic shielding, antenna design, intelligent reflecting surface and terahertz communication, etc. Metamaterials are generally considered physically as regular arrays of scatterers in space. By extending this concept, the basic physical model of metasurface, i.e. two-dimensional regular array of scatterers, can also be obtained.
[0003] The current design process of metasurface mainly includes: determining a series of units (meta-atoms) with similar structure and geometric composition; determining the amplitude and phase characteristics of different units; selecting and combining some units according to the target function. However, there are two problems to be solved in this process: first, the basic model of metamaterial is the scatterer model, whose unit structure is described as a scatterer parameter with a complexity of 36. However, the common design method currently only uses the surface current vector with a complexity of 3 or the phase with a complexity of 1 for design, which leads to the inability to generate arbitrary electric field distribution of metamaterial, and an all-complexity design method is urgently needed. Second, the current design method does not consider the mutual influence between metamaterial units, which leads to the inability to solve the non-local response, and a global design method is urgently needed. SUMMARY
[0004] The application proposes a non-local generative scatterer parameter design method, which mainly includes: first, a general model between scatterer parameters and generated electric field is established, and the mapping from the basic model of the metasurface to the basic characteristics of the electromagnetic field is realized; second, the interaction between multiple scatterers is explained and two forms of non-local response between scatterers are described; third, the reverse design of scattering characteristics is innovatively converted into a style transfer problem, and a method for generating scatterer parameter image-generation electric field image dataset is proposed; fourth, a generative adversarial network architecture with ResBlock is proposed, which realizes the overall design of any electric field distribution by optimizing all polarization rate parameters of the scatterer array.
[0005] The application proposes a non-local generative scatterer parameter design method for metasurface design. Unlike traditional metasurface design methods, this method can realize an ideal design process based on the basic model of the metasurface: first, obtain the scatterer parameters and arrangement according to the required function or electric field distribution; second, convert the scatterer model into a material reality. Compared with the current method, the above design method has three advantages: first, it does not require pre-design of metasurface units; second, it is not limited by geometric structure and material; third, it can realize any electric field distribution by simultaneously designing all degrees of freedom of scatterer parameters.
[0006] The method includes:
[0007] A mathematical model describing the mapping relationship between the scatterer model and the electric field distribution.
[0008] The scatterer model is the basic physical model of metamaterials, and the electric field distribution is the basic property of the electromagnetic field on the observation surface. Among them, the dipole emitter is located at the origin r'0=(0,0,0) of the Cartesian coordinate system, the scatterer array is rectangular, and N scatterers satisfy N=n 2 , located at It can be regarded as a source point of the electromagnetic field; the observation point plane is rectangular, and M observation points satisfy M=m 2 , and the position is It can be regarded as a field point of the electromagnetic field.
[0009] According to the Maxwell equations and dyadic Green's function, the electric field intensity at the observation point can be obtained as
[0010]
[0011] Among them, E(r j ) is the electric field intensity vector at r j , E0(r jis the electric field generated by the dipole emitter at r j , is the electric field generated by the dipole emitter at r j is the dyadic Green's function at r' i , is the curl of , and are the electric and magnetic polarizability tensors of the scatterer.
[0012] Preferably, the method further comprises:
[0013] An explanatory analysis of the two forms of non-local effects in the scatterer model.
[0014] The non-local effects of the scatterer model have the following forms: one is from r'1 to r'2 and back to r'1, which is represented as the square of the Green's function, i.e. multiplied by the electric field intensity or the magnetic field intensity at r'1; the other is from r'1 to r'2 and back to r'1, and finally back to r'2, which is scattered by r'2, which is represented as the cube of the Green's function, i.e. multiplied by the electric field intensity or the magnetic field intensity at r'2. This understanding can be extended to the general scatterer model, and the electric field component generated by any scatterer will include: electromagnetic waves that start from itself and are finally scattered by itself, and electromagnetic waves that start from itself and are finally scattered by other scatterers.
[0015] Preferably, the method further comprises:
[0016] A method for generating a scatterer parameter image-scattered electric field image data set.
[0017] Both the scatterer parameters and the electric field distribution have data dimensions suitable for imaging, and in this process, it is necessary to ensure: one-to-one correspondence between the scatterer parameters and the scattered electric field, randomness of the electric field distribution, physical realizability of the scatterer parameters, and spatial correlation when the scatterer parameters are converted into images.
[0018] Preferably, the method further comprises:
[0019] A generative adversarial network architecture with a ResBlock structure.
[0020] In the design of the generative network, ResBlock in ResNet is added, which solves the optimization difficulty problem and the gradient disappearance phenomenon in image generation.
[0021] The generative network is composed of 5 parts, including: a data input part, 2 down-sampling parts, 9 ResBlock parts, 2 up-sampling parts, and a data output part from input to output.
[0022] The adversarial network is composed of five parts. The first four parts include convolutional layers, BatchNorm layers and activation function Relu layers; the fifth part includes a convolutional layer and an activation function sigmoid layer. The convolution kernel size used in the adversarial network is 4.
[0023] A scatterer parameter global design method based on target electric field distribution according to the present application comprises the following steps:
[0024] Step 200, determine the number and arrangement of scatterers, determine the number and arrangement of observation points, and establish a mathematical model between the scatterer model and the scattering electric field distribution.
[0025] Step 210, randomly generate the polarizability parameter to obtain the generated electric field distribution of the observation surface.
[0026] Step 220, obtain the scatterer parameter image and the scattering electric field image, and obtain the data set.
[0027] Step 230, build a generative adversarial network and train it to obtain a trained model and save it.
[0028] Step 240, input the target electric field distribution to obtain the globally designed scatterer parameters.
[0029] Advantages
[0030] The present application proposes a scatterer parameter global design method based on a generative adversarial network. The forward mapping and inverse design between the scatterer properties and the generated electric field are realized, and the two forms of non-local effects in the scatterer model are explained. In this process, the scatterer property design problem is converted into a style migration problem in computer vision (computer vision, i.e. CV), and the complex polarizability tensor of the scatterer is optimized simultaneously to achieve any electric field distribution. The training results show that the model has good convergence, and the loss values of the generated network and the adversarial network after training are 0.0571 and 0.2220 respectively, and the average peak signal-to-noise ratio (peak signal-to-noise ratio, i.e. PSNR) between the electric field distribution corresponding to the generated scatterer parameters and the target electric field distribution reaches 103.1464dB. At the same time, this method can greatly reduce the time required for the design of metasurfaces. After testing, the average time required to generate 2304 scatterer parameters is only 0.1947 seconds. BRIEF DESCRIPTION OF DRAWINGS
[0031] Figure 1 is a flowchart for realizing a scatterer parameter global design method based on a target electric field distribution;
[0032] Figure 2is a global design method of scatterer parameters based on target electric field distribution, and the structure diagram, wherein the forward model is from the scatterer model to the generated electric field distribution, and the inverse algorithm is to solve the scatterer parameters according to the target electric field distribution;
[0033] Figure 3 is a mathematical model diagram of the mapping relationship between the scatterer model and the electric field distribution, and an explanatory analysis diagram of two forms of non-local effects in the scatterer model;
[0034] Figure 4 is a one-to-one correspondence diagram of scatterer parameter images and scatter electric field images, wherein the left image is a scatterer parameter image, the right image is a generated electric field image, and the upper left image is a local scatterer parameter diagram;
[0035] Figure 5 is a generative adversarial network architecture diagram with ResBlock structure, wherein the upper diagram is a generative network architecture diagram, the input of the generative network is an electric field distribution image, and the output is a scatterer parameter image, and the lower diagram is an adversarial network architecture diagram, the input of the adversarial network is a scatterer parameter image, and the output is a judgment result of whether the scatterer parameter image and the target electric field are one-to-one corresponding;
[0036] Figure 6 is the training result of the generative adversarial network, (a) and (b) are respectively the loss-training round number diagram and the learning rate / average peak signal-to-noise ratio diagram of the data set with 5000 pairs of images, (c) and (d) are respectively the loss-training round number diagram and the learning rate / average peak signal-to-noise ratio diagram of the data set with 10000 pairs of images, and in the legend, LD indicates the loss function value of the generative network, LG indicates the loss function value of the adversarial network, LR indicates the learning rate of the model, and PSNR indicates the average peak signal-to-noise ratio;
[0037] Figure 7 is a comparison of the electric field distribution images in the test set and the electric field distribution images corresponding to the scatterer parameters generated by taking them as target electric fields, (a) and (c) are respectively the electric field distribution images of the relatively complex pairs selected in the test set of the data sets with 5000 pairs and 10000 pairs of images, (b) and (d) are the electric field distribution images corresponding to the scatterer parameters generated by taking (a) and (c) as target electric fields, (e) and (f) are respectively the absolute error between (a) and (b) and (c) and (d);
[0038] Figure 8For the comparison of target electric field and generated electric field of single focus and four focus, (a) and (b) are the target electric field with a single focus and the generated electric field corresponding to the scatterer parameters generated by the generative adversarial network respectively, (c) and (d) are the target electric field with four focuses and the generated electric field corresponding to the scatterer parameters generated by the generative adversarial network respectively, (e), (f), (g) and (h) are the gray images of the above images, which are the images actually processed by the generative adversarial network;
[0039] Figure 9 are the normalized signal-to-noise ratio images of the target electric field and the generated electric field of single focus and four focus. (a) is the contrast image of the target electric field and the generated electric field signal-to-noise ratio image along the diagonal line of single focus, (b) is the contrast image of the target electric field and the generated electric field signal-to-noise ratio image along the diagonal line of four focus, wherein the solid line represents the trend of the target electric field signal-to-noise ratio image along the diagonal line, and the dotted line represents the trend of the generated electric field signal-to-noise ratio image along the diagonal line. DETAILED DESCRIPTION
[0040] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.
[0041] Figure 1 is a flow chart of a non-local generative scatterer parameter design method. For the required design of the metasurface function, the global scatterer parameter design method mainly includes: the number N and arrangement of scatterers in the scatterer model, the number M and arrangement of observation points are determined, and then a mathematical model from the scatterer model to the electric field distribution is established; the polarization rate parameters corresponding to each scatterer are randomly generated, and the physical realizability is verified, and then the modulus of the electric field intensity at all observation points is calculated to obtain the generated electric field distribution of the observation surface; the scatterer parameters and the generated electric field distribution are processed to obtain the one-to-one corresponding scatterer parameter image and the generated electric field image, which are sorted and classified to obtain the training set, the validation set and the test set respectively, taking 75% of the data set as the training set, 15% as the validation set and 5% as the test set; the near-field electric field distribution of the metasurface is obtained according to its function, and the target electric field distribution actually required is input into the training model to obtain the globally designed scatterer parameters.
[0042] Step 300, a method for establishing a mathematical model of the mapping relationship between the scatterer model and the electric field distribution.
[0043] The brief idea is as shown in Figure 3 , which can be expressed as follows: first, based on the dipole radiation theory and the scatterer model, the non-local scatter field is forward modeled. Secondly, the concept of observation point is further introduced, which is extended to the near-field two-dimensional or three-dimensional electric field distribution. Therefore, a mathematical model is established between the scatterer parameters of the source point and the electric field distribution of the field point. In space, this model can be considered as the mapping relationship between the scatterer array surface and the observation point surface.
[0044] In this modeling process, the array of scatterers needs to satisfy the following conditions to be consistent with the existing experimental results: first, the radius of the scatterer and the inverse of the wave number have the same order of magnitude; second, the distance between the scatterers is greater than or equal to 3 times the radius. This model first uses the Maxwell equations to describe the basic properties of electromagnetic waves and electromagnetic fields, which are expressed as
[0045]
[0046]
[0047] where E is the electric field intensity vector, H is the magnetic field intensity vector, J is the current density vector of the dipole emitter, and P and M are the polarizability and magnetizability, respectively.
[0048] By introducing the dyadic Green's function the electric field intensity and the magnetic field intensity considering the non-local response can be expressed as
[0049]
[0050] where E(r j ) is the electric field intensity vector at r j , H(r j ) is the magnetic field intensity vector at r j , and H0(r j ) is the magnetic field generated by the dipole emitter at r j .
[0051] Analyzing the above equation, it can be found that the electric field intensity and the magnetic field intensity at the observation point are considered to be the superposition of the radiation field generated by the dipole emitter and the scattering field generated by the scatterer. By observing the second term of the equation, it can be found that, in addition to the influence of each scatterer on the electric field at the observation point, the scattering field also needs to consider the interaction between different scatterers.
[0052] The radiation field E0(r) and H0(r) generated by the dipole emitter can be expressed as
[0053]
[0054]
[0055] The relationship between the radiation field and the scattering field at the observation point can be rewritten as the following matrix equation
[0056]
[0057] where R is the response matrix, which can be expressed as
[0058]
[0059]
[0060] It is worth noting that the response matrix R∈C 6N×6N is the most computationally complex part of the forward model since it is simultaneously for three spatial dimensions, for electric and magnetic fields, and for all N scatterers. At the same time, since this response matrix cannot be inverted using analytical methods, the inverse mapping from the generated electric field to the scatterer parameters cannot be analytically expressed.
[0061] Further, the modulus of the electric field strength at the observation point can be obtained as
[0062]
[0063] Since the above derivation process is independent of the scale and frequency, a unitless coordinate system can be used for modeling and calculation. First, select the wave number ξ = γk in the unitless system, and take γ = 1 / k for the convenience of calculation and expression. Considering the reciprocity of electric and magnetic fields, the spatial wave impedance in this unitless system is Z0 = 1.
[0064] Taking a simple system composed of two scatterers as an example, the electric field strength vector of each scatterer can be expressed as
[0065]
[0066]
[0067] The electric field strength vector of each scatterer can be expressed as
[0068]
[0069]
[0070] where γ is the determinant of the elements of the response matrix, which can be expressed as
[0071]
[0072] Analyzing a simple model composed of a dipole emitter and two scatterers, two conclusions can be obtained: first, due to the interaction between the dipole emitter and the scatterers and the non-local response between the scatterers, any E(r′ i ) or H(r′ i ) expression includes all elements in the response matrix and (E0(r1), H0(r1), E0(r2), H0(r2), …) TConsidering the model composed of N scatterers, the expression of the electric field intensity will contain 6(N-1)×6(N-1)×2N components; secondly, E(r′ i ) and H(r′ i ) has a completely symmetric structure with respect to i. This is because the response matrix and its own transpose matrix have a completely symmetric structure with respect to i. However, since different scatterers have different polarizability tensors and positions, this symmetric structure cannot be used to realize the inverse mapping from the generated electric field distribution to the scatterer parameters.
[0073] At the same time, by analyzing this simple model, we can get the following forms of non-local effects: First, the process from r′1 to r′2 and back to r′1 is expressed as the quadratic Green's function, that is, Multiply the electric field intensity or magnetic field intensity at r'1; secondly, from r'1 to r'2 and back to r'1, and finally back to r'2, scattered by r'2. This process is expressed as the cube of Green's function, that is, Multiply it by the electric or magnetic field strength at r′2.
[0074] The above understanding of nonlocal response can be extended to a model composed of N scatterers. The electric field components generated by any scatterer will include two forms: first, starting from itself and returning to itself. This process is represented by the even power of the Green's function, multiplied by its own electric field intensity or magnetic field intensity; second, starting from itself and emitted by other scatterers. This process is represented by the odd power of the Green's function, multiplied by the electric field intensity or magnetic field intensity at other scatterers.
[0075] Step 310: generating scatterer parameters and obtaining a method for generating an electric field distribution on an observation surface.
[0076] In order to ensure the randomness of the electric field distribution, and The real and imaginary parts of are generated by uniformly distributed random functions; in order to ensure its physical feasibility, it is necessary to verify whether the randomly generated data satisfies
[0077]
[0078] in Represents the unit tensor.
[0079] Based on randomly generated and By calculating the modulus of the electric field intensity at each observation point, the electric field intensity distribution on the observation surface can be obtained, thereby obtaining a one-to-one correspondence between the scatterer parameters and the scattered electric field distribution.
[0080] Step 320, obtain the scatterer parameter image and the scatter electric field image, and obtain the training set, the validation set and the test set.
[0081] The electric field distribution data is based on observation points arranged in a rectangular plane, which are uniformly arranged and naturally correspond to an m x m pixel image form, which is easy to convert into an image.
[0082] However, the scatterer parameters do not meet such an arrangement. With the development of the generative adversarial network, the generative network and the discriminative network are gradually realized by the convolutional neural network, which means the introduction of the convolution kernel. Therefore, when generating the scatterer parameter image, it is necessary to ensure its spatial correlation to meet the characteristics of the convolution kernel focusing on local features. The mapping relationship between each group of images in the data set is as shown in Figure 4 Each scatterer parameter is arranged according to the shape of the convolution kernel, and the position of the rectangle in the scatterer parameter image corresponds to the position of the scatterer in the array.
[0083] Based on the generated one-to-one corresponding scatterer parameter and scatter electric field distribution image, further, 75% of which is used as the training set, 15% as the validation set, and 5% as the test set, to obtain the data set for the generative adversarial network.
[0084] Step 330, method for building a specific structure generative adversarial network and training parameter setting. The basic structure of the generative network and the adversarial network in the generative adversarial network is as shown in Figure 5 .
[0085] The generative adversarial network is composed of a generative network and an adversarial network, and the training target of the generative network is to generate images that cannot be distinguished by the adversarial network, while the training target of the adversarial network is to distinguish as much as possible the false images generated by the generative network. The main idea is to find the mapping relationship between the random noise vector z and the output image y through the training process, which can be expressed as G: z→y, and the loss function in this process is expressed as
[0086]
[0087] Based on the generative adversarial network model, the main idea of the conditional generative adversarial network is to find the mapping relationship between the input data x and the random noise vector z and the output image y, which can be expressed as G: {x, z}→y, and the loss function can be expressed as
[0088]
[0089] The optimization target of the conditional generative adversarial network can be expressed as
[0090]
[0091] The generator with skips, Markovian discriminator and ResBlock in ResNet are introduced in the generative network used in the application. The main role of the generator with skips is to share some information between the input and output of the model, the main role of the Markovian discriminator is to make the network structure more sensitive to the high frequency part of the image, and the main role of the ResBlock is to reduce the optimization complexity in training and solve the gradient disappearance problem. In addition, the model introduces a regularization function to enhance the discriminant ability of the generative adversarial network, and the training process of the mapping relationship G:{x,z}→y can be optimized by learning the loss function. In summary, the model proposed in the application can be represented as
[0092]
[0093] The generative network includes 15 sub-networks of data input part, 2 down-sampling parts, 9 ResBlock parts, 2 up-sampling parts and data output part, and the specific composition is as follows: the data input part includes a ReflectionPad2d layer for symmetric expansion of data, a convolution layer with a convolution kernel size of 7, a BatchNorm layer for preventing gradient explosion or gradient disappearance phenomenon, and a Relu activation function layer; the down-sampling part includes a convolution layer with a convolution kernel size of 3, a BatchNorm layer, and a Relu activation function layer; the ResBlock layer includes two ReflectionPad2d layers and two convolution layers with a convolution kernel size of 3; the up-sampling part includes a convolution layer with a convolution kernel size of 3, a BatchNorm layer and a Relu activation function layer; the data output part includes a ReflectionPad2d layer, a convolution layer with a convolution kernel size of 7, and a Tanh activation function layer.
[0094] The generative adversarial network is composed of five sub-networks, and the basic structure is a convolutional neural network. The specific composition is as follows: the first four sub-networks include a convolution layer with a convolution kernel size of 4, a BatchNorm layer and an activation function Relu layer; the fifth sub-network includes a convolution layer with a convolution kernel size of 4 and an activation function sigmoid layer.
[0095] In the specific training process, the size of the Batchsize is determined by the storage space of the graphics card; the learning rate is decreased with the number of rounds, and the initial value is 0.0001 to 0.0002; the training rounds are 100 to 200 rounds, which can be adjusted according to the convergence; after testing, the data set including 20000 images has good convergence characteristics.
[0096] Step 340, obtaining the scatterer parameter method of global design.
[0097] The method of global design of scatterer parameters according to the existing model can be summarized as follows: first, according to the target function of the designed super surface, the mapping of the electric field distribution on the near-field observation point surface is obtained, which is specifically represented as the modulus of the electric field distribution vector; second, it is converted into an m x m pixel target electric field image as the input of the generated network in the saved generative adversarial network; third, using the generative adversarial network model, the scatterer parameter image is obtained as the output result.
[0098] As shown in Figure 8 To verify the ability of the method to generate any electric field as a whole, it is applied to the common near-field scene of focusing. The electric field distribution of single focus and four focus points is respectively taken as the target electric field of the model application. (a) and (b) are respectively the target electric field with a single focus point and the generated electric field corresponding to the scatterer parameters generated by the generative adversarial network, where the center coordinates of the focus point are (0, 0), and the side length of the rectangular focus region is (c) and (d) are respectively the target electric field with four focus points and the generated electric field corresponding to the scatterer parameters generated by the generative adversarial network, where the center coordinates of the focus points are (±2λ, ±2λ), and the side length of the rectangular focus region is (e), (f), (g) and (h) are the gray images of the above images, which are the images actually processed by the generative adversarial network. At the same time, in order to quantitatively evaluate the quality of the generated electric field, we define the signal-to-noise ratio (SNR) as the ratio of the electric field intensity of an observation point to the average electric field intensity of the entire observation surface, as shown in Figure 9 , where the solid line represents the change trend of the signal-to-noise ratio image of the target electric field along the diagonal line, and the dashed line represents the change trend of the signal-to-noise ratio image of the target electric field along the diagonal line.
[0099] Further, the scatterer parameter image is converted into the polarizability tensor in the scatterer model, thereby realizing the global design of the scatterer parameters based on the target electric field distribution.
[0100] Compared with the traditional method, the method has a significant time advantage. After testing, the time required to generate all 23040000 scatterer parameters of 10000 8x8 scatterer arrays is 1947 seconds, and the average time required to generate each 2304 scatterer parameters is 0.1947 seconds.
[0101] Those skilled in the art of the present technology should recognize that the above embodiments are only used to illustrate the present application, and are not intended to limit the present application, as long as the changes and modifications of the above embodiments fall within the scope of the present application.
Claims
1. A non-local generative scatterer parameter design method, characterized in that: include: Determine the number and arrangement of scatterers, determine the number and arrangement of observation points, and establish a mathematical model between the scatterer model and the scattered electric field distribution; randomly generate polarizability parameters to obtain the generated electric field distribution of the observation surface; convert the polarizability parameters and the generated electric field distribution into scatterer parameter images and generated electric field images to obtain a data set; build and train a generative adversarial network to obtain and save the training model; input the target electric field distribution to obtain the globally designed scatterer parameters; The mathematical model between the scatterer model and the generated electric field distribution is as follows: the dipole emitter is located at the origin r′0=(0,0,0) of the Cartesian coordinate system, the scatterer array is rectangular, and N scatterers satisfy N=n 2 ,lie in It can be regarded as the source point of the electromagnetic field; the observation point plane is a rectangle, and M observation points satisfy M=m 2 , whose location is It can be regarded as a field point of the electromagnetic field; According to Maxwell's equations and dyadic Green's function, the electric field intensity at the observation point can be obtained as Among them, E(r j ) is r j The electric field strength vector at the point, E0(r j ) is the dipole emitter at r j The electric field generated at For r j and r′ i The dyadic Green's function at for The curl of and is the electric susceptibility tensor and magnetic susceptibility tensor of a scatterer; The explanatory analysis of the two forms of non-local effects in the scatterer model is as follows: First, the process from r′1 to r′2 and back to r′1 is expressed as the quadratic Green's function, that is, Multiply it by the electric field intensity or magnetic field intensity at r'1; secondly, it goes from r'1 to r'2 and back to r'1, and is finally scattered by r'2. This process is expressed as the cube of Green's function, that is, Multiplied by the electric field intensity or magnetic field intensity at r′2; the above expression can be extended to the general scatterer model, and the electric field components generated by any scatterer will include: electromagnetic waves that start from itself and are finally scattered by itself, as well as electromagnetic waves that start from itself and are finally scattered by other scatterers.
2. The non-local generative scatterer parameter design method according to claim 1, characterized in that: A method for generating scatterer parameter images and scattered electric field image datasets converts randomly generated polarizability tensors and electric field distributions into scatterer parameter images and generated electric field distribution images. In this process, it is necessary to ensure: a one-to-one correspondence between scatterer parameters and generated electric fields, the randomness of the electric field distribution, the physical feasibility of the scatterer parameters, and the spatial correlation when the scatterer parameters are converted into images.
3. The non-local generative scatterer parameter design method according to claim 1, characterized in that: The generative adversarial network includes: adding ResBlock in ResNet to the design of the generative network to solve the optimization difficulties and gradient vanishing phenomenon in image generation; the generative network consists of 5 parts, from input to output, including: data input part, 2 downsampling parts, 9 ResBlock parts, 2 upsampling parts and data output part; the adversarial network consists of 5 convolutional neural sub-networks.