Method and system for constructing a hyper-fine grid full transmission array

Abstract

The application provides a kind of ultra-fine grid comprehensive transmission array construction method and system, comprising: step 1: pixel coding is carried out to array surface, and the entire super surface transmission array is divided into grid of preset scale;Step 2: horn is used for feeding in the transmission super surface, and the required waveform is obtained by controlling the feeding signal through the super surface;Step 3: data is collected by MATLAB and HFSS joint simulation;Step 4: the neural network structure is constructed, the collected data is characterized learning, and the training strategy of gradually increasing array resolution is adopted, the corresponding relationship between array fine structure and array transmission response is learned, and the structure code of super surface array is obtained according to the expected amplitude phase distribution on the observation surface.The application can realize the design of the maximum degree of freedom of super surface by high-resolution representation of array surface, fine grid division of the entire array, and then intelligent algorithm design.

Description

Technical Field

[0001] This invention relates to the field of metasurface transmission antenna technology, and more specifically, to a method and system for constructing an ultra-fine grid full transmission array. Background Technology

[0002] With the development of emerging technologies based on artificial intelligence (AI), machine learning (ML) and deep learning (DL) algorithms have received widespread attention for solving complex problems in electromagnetic fields such as remote sensing, inverse scattering, microwave device design, and antenna design.

[0003] Metasurfaces (MTS) are considered two-dimensional metamaterials, artificially engineered structures with unique electromagnetic properties not found in nature. MTS offer significant opportunities for electromagnetic wave control. In recent years, many researchers have employed machine learning algorithms such as artificial neural networks (ANNs), particle swarm optimization (PSO), and genetic algorithms (GA) to optimize and inversely design MTS antennas.

[0004] Antenna geometry optimization mainly falls into two categories. The first method uses machine learning (ML) algorithms to optimize the antenna scale using predefined patterns. For example, a meta-surface configured as three layers of equally sized square patches is optimized to the optimal scale, i.e., the side length of the patches. The meta-surface maintains its original square geometry during optimization. The second method optimizes the geometric pattern of the meta-surface. This is achieved by pixelating the planar antenna geometry into a binary matrix with 1b element states. By optimizing the binary matrix, antennas with higher geometric complexity are generated. Pixelated optimization design methods are introduced into MTS research. Although these MTSs are designed with pixelated geometry, their performance is similar to that of traditional design methods.

[0005] Patent document CN116247440A discloses an ultrathin transmissive / reflective dual-function metasurface antenna, belonging to the field of antenna technology. It includes an antenna system comprising a feed and an array of multiple transmissive / reflective elements; the feed is used to transmit or receive electromagnetic waves; and the transmissive / reflective elements are used to modulate the phase of the reflected wave and the phase of the transmitted wave. However, this patent cannot completely solve the existing technical problems. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for constructing an ultra-fine grid-based full-range transmission array.

[0007] The method for constructing an ultra-fine mesh full transmission array according to the present invention includes:

[0008] Step 1: Pixelate the array surface, dividing the entire metasurface transmission array into a grid of a preset size;

[0009] Step 2: Use a horn to feed the transmissive metasurface. The feed signal is modulated by the metasurface to obtain the desired waveform.

[0010] Step 3: Collect data using MATLAB and HFSS co-simulation;

[0011] Step 4: Construct a neural network structure, perform feature learning on the collected data, and adopt a training strategy of progressively increasing array resolution to learn the correspondence between the fine structure of the array and the transmission response of the array. Based on the expected amplitude and phase distribution on the observation surface, obtain the structural encoding of the metasurface array.

[0012] Preferably, the array coding matrix is ​​generated using MATLAB and data is collected through joint simulation using HFSS, and data collection and network training are performed at resolution scales of 8×8, 16×16, 32×32, 64×64, and 128×128, respectively.

[0013] Preferably, the neural network structure includes a generator and a discriminator. The required transmission response and random noise are fed into the generator, which then generates a matrix representing the pattern of the pixelated array. The required transmission response guides the generator to generate an array with such electromagnetic properties. When it comes to the discriminator, the input data are the transmission response and geometric matrix of the pixelated array.

[0014] Each layer in the generator and discriminator contains a batch normalization layer and a non-linear activation function;

[0015] If the actual geometry matrix of the real array is used, the actual transmission response obtained from the full-wave simulation is fed into the discriminator; if the generated geometry matrix is ​​used, the desired transmission response is fed into the discriminator.

[0016] The discriminator CNN is used to evaluate the input data of the discriminator. The output of the discriminator is a single scalar in the range [0, 1]. If the input data is a geometric matrix generated by the generator, the discriminator outputs 0; if the input data is obtained from the real geometric matrix, the output is 1.

[0017] Preferably, the generator includes a deconvolutional neural network with four layers, consisting of 256, 128, 64 and 1 output channels respectively;

[0018] The input data of the generator is a combination of two vectors: the desired transmission response and a random noise vector. The desired transmission response includes the amplitude response and the phase response, both of which are A×A matrices. The two matrices are reduced to one dimension and then concatenated. The random noise vector is randomly generated from a uniform distribution.

[0019] Preferably, the discriminator comprises five layers, each with 64, 128, 256, 512, and 1 output channel, respectively;

[0020] The input data of the discriminator is a combination of vectors and geometric matrices. When the geometric matrix is ​​obtained from the real element, the vector is sampled from the transmission response of the real element; when the geometric matrix is ​​obtained from the generated element, it is sampled from the expected transmission response.

[0021] The ultra-fine mesh full transmission array construction system provided by the present invention includes:

[0022] Module M1: Pixelates the surface, dividing the entire metasurface transmission array into a grid of a preset scale;

[0023] Module M2: A horn is used to feed the transmissive metasurface. The feed signal is modulated by the metasurface to obtain the desired waveform.

[0024] Module M3: Data acquisition is performed using a combination of MATLAB and HFSS simulations;

[0025] Module M4: Constructs a neural network structure, performs feature learning on the collected data, and adopts a training strategy of progressively increasing array resolution to learn the correspondence between the fine structure of the array and the transmission response of the array. Based on the desired amplitude and phase distribution on the observation surface, the structural encoding of the metasurface array is obtained.

[0026] Preferably, the array coding matrix is ​​generated using MATLAB and data is collected through joint simulation using HFSS, and data collection and network training are performed at resolution scales of 8×8, 16×16, 32×32, 64×64, and 128×128, respectively.

[0027] Preferably, the neural network structure includes a generator and a discriminator. The required transmission response and random noise are fed into the generator, which then generates a matrix representing the pattern of the pixelated array. The required transmission response guides the generator to generate an array with such electromagnetic properties. When it comes to the discriminator, the input data are the transmission response and geometric matrix of the pixelated array.

[0028] Each layer in the generator and discriminator contains a batch normalization layer and a non-linear activation function;

[0029] If the actual geometry matrix of the real array is used, the actual transmission response obtained from the full-wave simulation is fed into the discriminator; if the generated geometry matrix is ​​used, the desired transmission response is fed into the discriminator.

[0030] The discriminator CNN is used to evaluate the input data of the discriminator. The output of the discriminator is a single scalar in the range [0, 1]. If the input data is a geometric matrix generated by the generator, the discriminator outputs 0; if the input data is obtained from the real geometric matrix, the output is 1.

[0031] Preferably, the generator includes a deconvolutional neural network with four layers, consisting of 256, 128, 64 and 1 output channels respectively;

[0032] The input data of the generator is a combination of two vectors: the desired transmission response and a random noise vector. The desired transmission response includes the amplitude response and the phase response, both of which are A×A matrices. The two matrices are reduced to one dimension and then concatenated. The random noise vector is randomly generated from a uniform distribution.

[0033] Preferably, the discriminator comprises five layers, each with 64, 128, 256, 512, and 1 output channel, respectively;

[0034] The input data of the discriminator is a combination of vectors and geometric matrices. When the geometric matrix is ​​obtained from the real element, the vector is sampled from the transmission response of the real element; when the geometric matrix is ​​obtained from the generated element, it is sampled from the expected transmission response.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] This invention proposes a novel metasurface design approach, opening up a new path for research combining deep learning and metasurface design. Traditional metasurface inverse design based on intelligent algorithms only involves meshing the units. This invention, by repeatedly increasing the resolution of the pixelated array, finely divides the entire array into meshes, and then uses intelligent algorithms for design. By utilizing a training strategy that progressively increases the resolution, it is possible to achieve the design of the metasurface with the maximum degree of freedom, without the need for metasurface unit division. Attached Figure Description

[0037] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0038] Figure 1 A schematic diagram of the array surface coding structure for an ultra-fine mesh full transmission array based on a progressively growing convolutional generative adversarial network is provided for this invention.

[0039] Figure 2 A schematic diagram of a method for designing an ultra-fine mesh full-spectrum transmission array based on a progressively growing convolutional generative adversarial network provided by the present invention;

[0040] Figure 3a and Figure 3b A schematic diagram of the DCGAN model network structure constructed based on a progressively growing convolutional generative adversarial network with an ultra-fine mesh full-transmission array design, provided by the present invention.

[0041] Figure 4 (a) and (b) in the figure are schematic diagrams of a progressive growth strategy for designing an ultra-fine mesh full-transmission array based on a progressively growing convolutional generative adversarial network provided by the present invention. Detailed Implementation

[0042] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0043] Example

[0044] like Figure 1 and Figure 2 This invention provides a comprehensive transmissive array design based on a progressively growing convolutional generative adversarial network (PGAN), including overall array design, neural network structure, and training strategy. In the overall design of the metasurface array, the entire transmissive array is divided into a grid of a certain scale, equivalent to pixelating the array. The neural network structure combines a deep convolutional PGAN with a progressively growing PGAN, introducing this deep learning algorithm into the metasurface array design, providing an alternative approach. Furthermore, by repeatedly collecting data from arrays at different resolutions and using a progressively growing data resolution training strategy, the network can master the design of high-resolution arrays, i.e., fine array design.

[0045] The transmissive metasurface is fed by a horn, and the feeding signal is modulated by the metasurface to obtain the desired waveform. Data is acquired through joint simulation using MATLAB and HFSS, and a convolutional generative adversarial network (GAN) is constructed to learn features. A training strategy with progressively increasing array resolution is employed to learn the correspondence between the fine structure of the array and its transmission response. Based on the desired amplitude and phase distribution on the observation surface, the structural encoding of the metasurface array is directly obtained, achieving one-step metasurface array design.

[0046] Specifically, the overall array design directly targets the entire metasurface array, dividing the array into uniform grids of different sizes for encoding design, skipping unit design. This overall array design is combined with deep learning, employing a novel neural network structure in both deep learning and metasurface design. The neural network used in the overall array design draws inspiration from deep convolutional generative adversarial networks (GANs) and progressively growing GANs, combining the former's network structure with the latter's training strategy. The overall array design considers a frequency of 12.5 GHz and a metasurface side length of 5 wavelengths (λ), i.e., 120 mm. The training strategy requires collecting array data at different resolutions separately, providing data of different resolutions at different stages of network training, and simultaneously changing the network structure. The neural network structure consists of a generator and a discriminator. The required transmission response and random noise are fed into the generator. The generator then generates a matrix representing the pattern of the pixelated array. The required transmission response guides the generator to generate an array with this electromagnetic property. When it comes to the discriminator, the input data is the transmission response and geometric matrix of the pixelated array. If the geometry matrix of the real frontal array is used, the actual transmission response obtained from the full-wave simulation is fed into the discriminator. Otherwise, if the generated geometry matrix is ​​used, the desired transmission response is fed into the discriminator. The discriminator's input data is then evaluated using a discriminator CNN. The discriminator's output is a single scalar in the range [0, 1]. If the input data is a geometry matrix generated by the generator, the discriminator should output 0. Furthermore, if the input data is obtained from a real geometry matrix, it should output 1.

[0047] The generator network consists of a deconvolutional neural network with four layers, each with 256, 128, 64, and 1 output channel respectively. The number 1 indicates the layer number in the multilayer array. The generator network's input data is a combination of two vectors: the desired transmission response (amplitude and phase) and random noise, with a dimension of M. The desired transmission response includes the amplitude response and phase response, each an A×A matrix, which are then one-dimensionalized into vectors and concatenated. The random noise vector is randomly generated from a uniform distribution. The discriminator network's input data is a combination of vectors and a geometric matrix. When the geometric matrix is ​​obtained from the real element, the vector is sampled from the transmission response of the real element. When the geometric matrix is ​​obtained from the generated element, it is sampled from the desired transmission response.

[0048] During training, parameters including M, the desired frequency range, and the desired transmission response are adjustable, and N also varies accordingly with the incremental growth strategy, where N represents the current array resolution. These parameters are chosen to ensure that all training data is sampled under the same conditions. Figure 3 is a schematic diagram of the constructed network architecture. Real or generated geometric matrices are fed into a CNN discriminator model that generates a single scalar. The discriminator consists of five layers with 64, 128, 256, 512, and 1 output channel, respectively. Each layer in the generator and discriminator networks contains a batch normalization layer and a non-linear activation function (Leaky ReLU and Sigmoid).

[0049] In addition, the network structure also includes the changes generated in the generator output layer and the discriminator input layer during the training process. The overall array design (1) uses MATLAB to generate the array encoding matrix and HFSS to jointly simulate and collect data, and performs data collection and model training at resolution scales of 8×8, 16×16, 32×32, 64×64, and 128×128 respectively.

[0050] In Figure (a) 4, assuming a switch from 8×8 resolution to 16×16 resolution, after training on the 8×8 resolution for a sufficient number of iterations (a), another transposed convolution is introduced in the generator (G), and another convolution is introduced in the discriminator (D), making the "interface" between G and D 16×16. There are two paths to generate the 16×16 layer: (1-α) multiplication by a layer that simply scales using nearest-neighbor interpolation, which has no trained parameters and is relatively straightforward; and (α) multiplication by the output layer of an additional transposed convolution, which requires training but ultimately performs better. These two are connected to form a new 16×16 generated image, with α scaling linearly from 0 to 1. When α reaches 1, the nearest-neighbor interpolation from 8×8 will be completely zero. This smooth transition mechanism greatly stabilizes the network architecture, giving the system time to adapt to higher resolutions.

[0051] Under this strategy, after training to a certain extent at low resolution, the algorithm jumps to a higher resolution, but not immediately. Instead, it smoothly adds new layers with higher resolution through a parameter α (which scales linearly from 0 to 1). α affects the utilization of both the old, scaled-up layers and the newly generated, larger layers. In the discriminator D part, it simply scales down to half its original size and then smoothly injects the trained layers for discrimination, as shown below. Figure 4 As shown in (b), if you are confident in this new layer, keep it at 16×16, and then prepare to grow it again after properly training the 16×16 resolution layer.

[0052] In progressively growing convolutional generative adversarial networks (GANs), multiple data acquisitions of units at different resolutions are performed. Simultaneously, model training is divided into multiple stages: first, the model is trained at low resolution, and then the unit resolution is gradually increased. Corresponding network layers are added to the generator and discriminator. This design allows for more detailed design of the network's units.

[0053] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0054] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A method for constructing an ultra-fine mesh full-range transmission array, characterized in that, include: Step 1: Pixelate the array surface, dividing the entire metasurface transmission array into a grid of a preset size; Step 2: Use a horn to feed the transmissive metasurface. The feed signal is modulated by the metasurface to obtain the desired waveform. Step 3: Collect data using MATLAB and HFSS co-simulation; Step 4: Construct a neural network structure, perform feature learning on the collected data, and adopt a training strategy of progressively increasing array resolution to learn the correspondence between the fine structure of the array and the transmission response of the array. Based on the expected amplitude and phase distribution on the observation surface, obtain the structural encoding of the metasurface array. The neural network structure includes a generator and a discriminator. The required transmission response and random noise are fed into the generator, which then generates a matrix representing the pattern of the pixelated array. The required transmission response guides the generator to generate an array with such electromagnetic properties. When it comes to the discriminator, the input data is the transmission response and geometric matrix of the pixelated array. Each layer in the generator and discriminator contains a batch normalization layer and a non-linear activation function; If the geometry matrix of the real front surface is used, the actual transmission response obtained from the full-wave simulation will be fed into the discriminator; If the generated geometry matrix is ​​used, the desired transmission response is fed into the discriminator; The discriminator CNN is used to evaluate the input data of the discriminator. The output of the discriminator is a single scalar in the range [0, 1]. If the input data is a geometric matrix generated by the generator, the discriminator outputs 0; if the input data is obtained from the real geometric matrix, the output is 1. The generator includes a deconvolutional neural network with four layers, consisting of 256, 128, 64, and 1 output channels, respectively. The input data of the generator is a combination of two vectors: the desired transmission response and a random noise vector. The desired transmission response includes the amplitude response and the phase response, both of which are A×A matrices. The two matrices are transformed into vectors and concatenated. The random noise vector is randomly generated from a uniform distribution; The discriminator comprises five layers, with 64, 128, 256, 512, and 1 output channel respectively; The input data of the discriminator is a combination of vectors and geometric matrices. When the geometric matrix is ​​obtained from the real element, the vector is sampled from the transmission response of the real element. When the geometric matrix is ​​obtained from the generated meta-cells, sampling is performed from the desired transmission response.

2. The method for constructing an ultra-fine mesh full-range transmission array according to claim 1, characterized in that, The array coding matrix was generated using MATLAB and data was collected through joint simulation using HFSS. Data acquisition and network training were performed at resolution scales of 8×8, 16×16, 32×32, 64×64, and 128×128, respectively.

3. A system for constructing an ultra-fine mesh full-range transmission array, characterized in that, include: Module M1: Pixelates the surface, dividing the entire metasurface transmission array into a grid of a preset scale; Module M2: A horn is used to feed the transmissive metasurface. The feed signal is modulated by the metasurface to obtain the desired waveform. Module M3: Data acquisition is performed using a combination of MATLAB and HFSS simulations; Module M4: Constructs a neural network structure, performs feature learning on the collected data, and adopts a training strategy of progressively increasing array resolution to learn the correspondence between the fine structure of the array and the transmission response of the array. Based on the desired amplitude and phase distribution on the observation surface, the structural encoding of the metasurface array is obtained. The neural network structure includes a generator and a discriminator. The required transmission response and random noise are fed into the generator, which then generates a matrix representing the pattern of the pixelated array. The required transmission response guides the generator to generate an array with such electromagnetic properties. When it comes to the discriminator, the input data is the transmission response and geometric matrix of the pixelated array. Each layer in the generator and discriminator contains a batch normalization layer and a non-linear activation function; If the geometry matrix of the real front surface is used, the actual transmission response obtained from the full-wave simulation will be fed into the discriminator; If the generated geometry matrix is ​​used, the desired transmission response is fed into the discriminator; The discriminator CNN is used to evaluate the input data of the discriminator. The output of the discriminator is a single scalar in the range [0, 1]. If the input data is a geometric matrix generated by the generator, the discriminator outputs 0; if the input data is obtained from the real geometric matrix, the output is 1. The generator includes a deconvolutional neural network with four layers, consisting of 256, 128, 64, and 1 output channels, respectively. The input data of the generator is a combination of two vectors: the desired transmission response and a random noise vector. The desired transmission response includes the amplitude response and the phase response, both of which are A×A matrices. The two matrices are transformed into vectors and concatenated. The random noise vector is randomly generated from a uniform distribution; The discriminator comprises five layers, with 64, 128, 256, 512, and 1 output channel respectively; The input data of the discriminator is a combination of vectors and geometric matrices. When the geometric matrix is ​​obtained from the real element, the vector is sampled from the transmission response of the real element. When the geometric matrix is ​​obtained from the generated meta-cells, sampling is performed from the desired transmission response.

4. The ultra-fine mesh full-range transmission array construction system according to claim 3, characterized in that, The array coding matrix was generated using MATLAB and data was collected through joint simulation using HFSS. Data acquisition and network training were performed at resolution scales of 8×8, 16×16, 32×32, 64×64, and 128×128, respectively.

Images