A method for predicting scattering parameters of meta-atoms and a method for constructing meta-devices

Abstract

The present invention discloses a method for predicting scattering parameters of meta-atoms and a method for constructing meta-devices, which relates to the field of meta-atoms. Among them, the scattering parameter prediction method can effectively decompose the meta-atom design and extract its spatial and physical features by constructing an initial network model including a splitting module, a first channel, a second channel, a stacking module, a comprehensive feature channel and a calculation module. At the same time, by training the initial network model with meta-atom design samples with multiple different design structures and design parameters, the generalization ability of the model is greatly improved. The method of the present invention can accurately predict the real and imaginary parts of the transmission coefficient corresponding to the meta-atom in a short time, and further calculate the scattering parameters, greatly improving the efficiency and accuracy of the meta-atom scattering parameter prediction.

Description

Technical Field

[0001] The present invention relates to the field of meta-atoms, and particularly to a method for predicting scattering parameters for meta-atoms and a method for constructing a meta-device. Background Art

[0002] In the design process of metasurfaces and meta-atoms, modeling and characterization tools play a crucial role. To meet the stringent requirements of next-generation components for freeform and multifunctionality, researchers have been working on developing more reliable and efficient modeling tools. An early approach was to simplify the simulation process of metamaterials by establishing analytical effective medium models, such as the Lewin model and the GEM model. However, these models are only applicable to long-wavelength approximations under specific conditions and can only handle metamaterials with microsphere shapes. Another commonly used method relies on iterative full-wave numerical simulation techniques, including the finite element method (FEM), the finite-difference time-domain method (FDTD), and the finite integration technique (FIT). Although this method can provide high-precision results, its high computational cost limits its practicality.

[0003] Recently, data-driven modeling tools based on deep neural networks (DNN) have brought new breakthroughs to this field. Compared with traditional methods, this new type of tool has not only proven its accuracy and timeliness but also demonstrated significant advantages in predicting the optical response of nanophotonic structures. For example, DNN models using fully connected layers (FCL) have been successfully applied to predict the amplitude response of cylindrical, elliptical cylinder, spherical, and rod-shaped meta-atoms. More importantly, some studies have also shown that DNN can be used to predict phase responses, which are crucial for most optical applications that require precise control of incident light. After sufficient training, these models can generate electromagnetic responses within milliseconds, thus greatly accelerating the design process of meta-atoms / metasurfaces.

[0004] However, existing deep neural network models based on FCL also have certain limitations. First, they are mainly applicable to meta-atom structures that can be simply described by 3 to 5 parameters, which limits the possibility of achieving high efficiency and wide phase coverage, especially in the application of composite metasurfaces. Second, these models usually work in a very limited design space, where the key design parameters involved (such as lattice size, meta-atom thickness, and material properties) are fixed. Once these parameters change, it is necessary to re-collect data and re-train the model, which is a time-consuming process. Due to the existence of these weaknesses, the current deep neural network-based methods are limited in generalization ability, affecting their potential for widespread application. Summary of the Invention

[0005] To solve the above technical problems, the present invention proposes a method for predicting scattering parameters for meta-atoms, including:

[0006] Construct an initial network model including a splitting module, a first channel, a second channel, a stacking module, a comprehensive feature channel, and a computing module; where:

[0007] The splitting module is used to decompose the meta-atom design into a two-dimensional image tensor and a one-dimensional attribute tensor; the meta-atom design includes a design structure and design parameters; the design structure corresponds to the two-dimensional image tensor, and the design parameters correspond to the one-dimensional attribute tensor;

[0008] The first channel is used to extract the spatial features of the two-dimensional image tensor through convolution and pooling operations to obtain an image feature vector;

[0009] The second channel is used to extract the physical features of the one-dimensional attribute tensor through non-linear transformation and spatial tiling operations to obtain an attribute feature vector;

[0010] The stacking module is used to stack the image feature vector and the attribute feature vector to obtain a concatenated feature;

[0011] The comprehensive feature channel is used to process the concatenated feature through convolution and pooling operations to obtain an electromagnetic response feature;

[0012] The computing module is used to obtain the real part and the imaginary part of the transmission coefficient corresponding to the meta-atom design according to the electromagnetic response feature, and calculate the scattering parameters based on the real part and the imaginary part of the transmission coefficient; where the scattering parameters include phase and amplitude;

[0013] Generate multiple meta-atom designs with different design structures and design parameters, calculate the scattering parameters of the meta-atom designs, and label the corresponding meta-atom designs with the scattering parameters to obtain meta-atom design samples, and form a data set through the meta-atom design samples; the design parameters include: refractive index, thickness, and lattice size;

[0014] Train the initial network model through the data set to obtain a scattering parameter prediction model; predict the scattering parameters of the meta-atom design to be measured through the scattering parameter prediction model.

[0015] Further, the first channel includes:

[0016] A first convolutional layer, used to perform the first convolutional calculation on the two-dimensional image tensor to obtain the first spatial feature;

[0017] A first max-pooling layer, used to perform max-pooling operation on the first spatial feature to obtain a pooled feature;

[0018] A second convolutional layer, used to perform the second convolutional calculation on the pooled feature to obtain the second spatial feature;

[0019] The second max pooling layer is used to perform a max pooling operation on the second spatial feature to obtain an image feature vector.

[0020] Further, the second channel includes:

[0021] A non-linear transformation layer is used to perform a non-linear transformation on the one-dimensional attribute tensor to obtain a transformed feature;

[0022] A spatial tiling layer is used to perform a spatial tiling operation on the transformed feature to obtain an attribute feature vector.

[0023] Further, the comprehensive feature channel includes, connected in sequence:

[0024] A third convolutional layer is used to perform a convolution operation on the concatenated feature to obtain an initial convolutional feature;

[0025] A third max pooling layer is used to perform a max pooling operation on the initial convolutional feature to obtain an initial pooling feature;

[0026] A fourth max pooling layer is used to perform a max pooling operation on the initial pooling feature again to obtain a target pooling feature;

[0027] A fourth convolutional layer is used to perform a convolution calculation on the target pooling feature and flatten the result of the convolution calculation to obtain a target convolutional feature;

[0028] A first fully connected layer is used to perform a fully connected operation on the target convolutional feature to obtain an initial fully connected feature;

[0029] A second fully connected layer is used to transform the initial fully connected feature to obtain an electromagnetic response feature.

[0030] Further, in the comprehensive feature channel, the third convolutional layer, the third max pooling layer, the fourth max pooling layer, the fourth convolutional layer and the first fully connected layer all include a batch normalization unit and a ReLU activation function;

[0031] The batch normalization unit is used to perform batch normalization on the features obtained after performing the corresponding operations on their respective layers;

[0032] The ReLU activation function is used to perform a non-linear transformation on the result of the batch normalization process to finally obtain the output feature corresponding to this layer; the output feature is the initial pooling feature, the target pooling feature, the target convolutional feature or the initial fully connected feature.

[0033] Further, the generation of the meta-atom designs with multiple different design structures and design parameters is specifically:

[0034] Select a square canvas with a preset size as the design space for each meta-atom;

[0035] For each meta - atom, design a preset number of rectangular bars in the upper - left quadrant of the canvas based on randomly selected parameters;

[0036] The randomly selected parameters include: the number, size, and position of the rectangular bars; each rectangular bar represents a high - refractive - index material of a set type, and the rectangular bars correspond one - to - one with the types of high - refractive - index materials;

[0037] Mirror - copy the design in the upper - left quadrant along the x - axis and y - axis to form the complete design of each meta - atom;

[0038] By adjusting the randomly selected parameters and repeating the above steps, generate multiple meta - atom designs with different design structures and design parameters.

[0039] Furthermore, calculating the scattering parameters of the meta - atom design specifically includes:

[0040] Set boundary conditions: Use unit - cell boundary conditions to simulate the electromagnetic response of the meta - atom design in a square lattice structure, and apply open - boundary conditions in the direction perpendicular to the plane of the meta - atom design, i.e., the z - direction, to ensure that the wavefront can propagate freely;

[0041] Define the illumination method: Use an x - polarized plane - wave to illuminate the meta - atom design from the substrate side to simulate the incident light situation in an actual application scenario;

[0042] Determine the frequency range: Calculate the electromagnetic response of the meta - atom design in the frequency band of 30 to 60 THz, including transmittance and phase - shift information;

[0043] Based on the set boundary conditions, illumination method, and frequency range, calculate the transmittance and phase - shift of the meta - atom design through a full - wave electromagnetic simulation tool;

[0044] According to the obtained transmittance and phase - shift data, calculate the phase and amplitude of the meta - atom design.

[0045] The present invention also proposes a method for constructing a meta - device, including:

[0046] S1: Design a metasurface phase mask and decompose the performance requirements of the metasurface phase mask into those of each cell; one cell of the metasurface phase mask represents one meta - atom;

[0047] S2: Assign a serial number to the cell, and the initial value of the serial number is equal to 1;

[0048] S3: Input the performance requirements of the cell corresponding to the current serial number into a pre - trained GAN model;

[0049] S4: Generate a set of meta - atom designs through the GAN model; the set of meta - atom designs includes multiple meta - atom designs;

[0050] S5: Sequentially input the meta-atom designs in the meta-atom design set into the scattering parameter prediction model as described above for prediction; based on the model prediction results corresponding to each meta-atom design in the meta-atom design set, obtain the optimal meta-atom design for the corresponding cell.

[0051] S6: Determine whether the current serial number is greater than or equal to the maximum serial number of the cell. If not, increment the serial number by 1 and return to step S3; if so, proceed to the next step.

[0052] S7: Assemble the optimal meta-atom designs of each cell into a meta-device.

[0053] Further, the step S5 specifically includes:

[0054] S50: Set the target number of iterations and initialize the number of iterations to 1.

[0055] S51: Sequentially input the meta-atom designs in the meta-atom design set into the scattering parameter prediction model as described above for prediction.

[0056] S52: Based on the model prediction results of each meta-atom design, calculate its corresponding key performance indicators.

[0057] S53: Select the best meta-atom design in the meta-atom design set as the current optimal design based on the key performance indicators; determine whether the current number of iterations is greater than or equal to the target number of iterations. If not, increment the number of iterations by 1 and proceed to the next step; if so, jump to step S55.

[0058] S54: Feed the current optimal design back to the GAN model. The GAN model uses the current optimal design as a reference design to regenerate a new meta-atom design set and return to step S51.

[0059] S55: Then set the current optimal design as the optimal meta-atom design and terminate the iteration.

[0060] Further, the key performance indicators include transmission efficiency and phase shift accuracy.

[0061] The beneficial effects of the embodiments of the present invention include:

[0062] (1) By constructing an initial network model that includes a splitting module, a first channel, a second channel, a stacking module, a comprehensive feature channel, and a calculation module, the present invention can effectively decompose the meta-atom design and extract its spatial and physical features. At the same time, by training the initial network model with meta-atom design samples of multiple different design structures and design parameters, the generalization ability of the model is greatly improved. The method of the present invention can accurately predict the real and imaginary parts of the transmission coefficient corresponding to the meta-atom design in a short time, and further calculate the scattering parameters, greatly improving the efficiency and accuracy of meta-atom scattering parameter prediction.

[0063] (2) In the present invention, a square canvas of a preset size is selected as the design space for each meta-atom, and the meta-atom is allowed to be designed based on randomly selected parameters (such as the number, size, and position of rectangular bars) in the upper left quadrant, so that the design is no longer limited to a fixed set of parameters. This greatly increases the diversity and complexity of the design, which helps to achieve high efficiency and wide phase coverage. At the same time, the design in the upper left quadrant is mirror-copied along the x-axis and y-axis to form a complete meta-atom design. This method not only simplifies the design process but also ensures the consistency and symmetry of the design, which helps to improve the stability of the electromagnetic response.

[0064] (3) In the present invention, since each design can generate a unique meta-atom structure according to different random parameters, this diversified design method can effectively expand the working range of the model, making it no longer limited to a specific set of design parameters. Even in the face of changing design parameters, it can quickly adapt and generate corresponding designs, reducing the need to re-collect data and re-train the model.

[0065] (4) The present invention can select the optimal meta-atom design by using a pre-trained GAN model to generate multiple meta-atom designs and using the above-mentioned scattering parameter prediction model to evaluate the performance of each design. This process ensures that each cell can meet specific performance requirements, thereby improving the design and manufacturing accuracy of the entire meta-device.

[0066] (5) In step S5 of the present invention, an iterative optimization mechanism is introduced. By continuously feeding the current optimal design back to the GAN model to generate a new set of meta-atom designs and repeating the evaluation and selection process until a predetermined number of iterations is reached. This method effectively ensures the superiority of the finally selected meta-atom design in key performance indicators such as transmission efficiency and phase shift accuracy, achieving the maximization of the performance of the meta-device.

[0067] Details of one or more embodiments of the present invention are set forth in the following drawings and description to make the other features, objects, and advantages of the present invention more concise and understandable. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the accompanying drawings required in the description of the embodiments or the prior art. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other embodiments can be obtained based on these drawings.

[0069] Figure 1 is the structural diagram of the initial network model of the embodiment of the present invention;

[0070] Figure 2 is the flowchart of the construction method of a meta-device in the embodiment of the present invention;

[0071] Figure 3 is the design structure diagram of the meta-atom design in the embodiment of the present invention;

[0072] Figure 4 is the electric field distribution of the transmissive metasurface lens under different iteration times in the embodiment of the present invention. Detailed implementation manners

[0073] The following will describe the embodiments of this embodiment in more detail with reference to the accompanying drawings. Although some embodiments of this embodiment are shown in the accompanying drawings, it should be understood that this embodiment can be implemented in various forms and should not be construed as limited to the embodiments described herein. On the contrary, these embodiments are provided to more thoroughly and completely understand this embodiment. It should be understood that the accompanying drawings and embodiments of this embodiment are only for exemplary purposes and are not used to limit the protection scope of this embodiment.

[0074] Embodiment 1

[0075] In order to improve the generalization ability of the model and at the same time improve the efficiency and accuracy of the scattering parameter prediction of the meta-atom, the embodiment of the present invention proposes a method for predicting the scattering parameters of the meta-atom, including:

[0076] Construct an initial network model as shown in Figure 1 including a splitting module, a first channel, a second channel, a stacking module, a comprehensive feature channel, and a calculation module; where:

[0077] The splitting module is used to decompose the meta-atom design into a two-dimensional image tensor (size: 32×32) and a one-dimensional attribute tensor (size: 1×3); the meta-atom design includes a design structure and design parameters; the design structure corresponds to the two-dimensional image tensor, and the design parameters correspond to the one-dimensional attribute tensor;

[0078] The first channel is used to extract the spatial features of the two-dimensional image tensor through convolution and pooling operations to obtain an image feature vector;

[0079] The first channel includes:

[0080] A first convolutional layer (with a kernel size of 3×3 and 32 output channels), which is used to perform the first convolutional calculation on the two-dimensional image tensor to obtain the first spatial feature;

[0081] A first max pooling layer, which is used to perform a max pooling operation on the first spatial feature to obtain a pooled feature (size: 16×16×32);

[0082] A second convolutional layer (with a kernel size of 3×3 and 64 output channels), which is used to perform the second convolutional calculation on the pooled feature to obtain the second spatial feature;

[0083] A second max pooling layer, which is used to perform a max pooling operation on the second spatial feature to obtain an image feature vector (size: 8×8×64).

[0084] The second channel is used to extract the physical features of the one-dimensional attribute tensor through non-linear transformation and spatial tiling operations to obtain an attribute feature vector;

[0085] The second channel includes:

[0086] A non-linear transformation layer (with a kernel size of 64×3×3 and 32 output channels), which is used to perform a non-linear transformation on the one-dimensional attribute tensor to obtain a transformed feature;

[0087] A spatial tiling layer, which is used to perform a spatial tiling operation on the transformed feature to obtain an attribute feature vector (size: 8×8×64).

[0088] The stacking module is used to stack the image feature vector and the attribute feature vector to obtain a concatenated feature (size: 8×8×128);

[0089] The comprehensive feature channel is used to process the concatenated feature through convolutional and pooling operations to obtain an electromagnetic response feature;

[0090] The comprehensive feature channel includes the following connected in sequence:

[0091] A third convolutional layer (with a kernel size of 3×3 and 128 output channels), which is used to perform a convolutional operation on the concatenated feature to obtain an initial convolutional feature;

[0092] A third max pooling layer, which is used to perform a max pooling operation on the initial convolutional feature to obtain an initial pooled feature (size: 4×4×128);

[0093] A fourth max pooling layer, which is used to perform a max pooling operation on the initial pooled feature again to obtain a target pooled feature (size: 2×2×256);

[0094] The fourth convolutional layer (with a convolutional kernel size of 3×3 and 256 output channels) is used to perform convolutional calculations on the target pooled features, and flatten the result obtained from this convolutional calculation (the size of the convolutional result remains 2×2×256) to obtain the target convolutional features (size: 1×1024);

[0095] The first fully connected layer is used to perform a fully connected operation on the target convolutional features to obtain the initial fully connected features (size: 1×256);

[0096] The second fully connected layer is used to transform the initial fully connected features to obtain the electromagnetic response features (size: 1×51).

[0097] In the comprehensive feature channel, the third convolutional layer, the third max pooling layer, the fourth max pooling layer, the fourth convolutional layer, and the first fully connected layer all include batch normalization units and ReLU activation functions;

[0098] The batch normalization unit is used to perform batch normalization on the features obtained after performing corresponding operations on their respective layers;

[0099] The ReLU activation function is used to perform a non-linear transformation on the result of the batch normalization process to finally obtain the output features corresponding to this layer; the output features are the initial pooled features, the target pooled features, the target convolutional features, or the initial fully connected features.

[0100] The calculation module is used to obtain the real part and the imaginary part of the transmission coefficient corresponding to the meta-atom design according to the electromagnetic response features, and calculate the scattering parameters based on the real part and the imaginary part of the transmission coefficient; wherein, the scattering parameters include phase and amplitude;

[0101] Use the "pin-drop method" to generate meta-atom designs with multiple different design structures and design parameters, calculate the scattering parameters of the meta-atom designs, and label the corresponding meta-atom designs with the scattering parameters to obtain meta-atom design samples, and form a data set through the meta-atom design samples; the design parameters include: refractive index, thickness, and lattice size;

[0102] The generation of meta-atom designs with multiple different design structures and design parameters is specifically as follows:

[0103] Select a square canvas with a preset size (32x32 pixels) as the design space for each meta-atom;

[0104] For each meta-atom, design a preset number (3 to 7) of rectangular bars in the upper left quadrant of the canvas based on randomly selected parameters, so that each meta-atom has a unique initial design; the minimum resolution of each rectangular bar is 1 pixel;

[0105] The randomly selected parameters include: the number, size, and position of the rectangular bars; each rectangular bar represents a high refractive index material of a set type, and the rectangular bars correspond one-to-one with the types of high refractive index materials;

[0106] Mirror-copy the design in the upper left quadrant along the x-axis and y-axis to form the complete design of each meta-atom;

[0107] By adjusting the randomly selected parameters and repeating the above steps, multiple meta-atom designs with different design structures and design parameters are generated.

[0108] In total, 53,000 meta-atom designs with different shapes, refractive indices, thicknesses, and lattice sizes are generated in this embodiment. In this embodiment Figure 3 Four different meta-atom designs are shown. The red lines divide the canvas into four quadrants, and different colors (purple, green, blue) represent different rectangular bars.

[0109] In the present invention, by selecting a square canvas of a preset size as the design space for each meta-atom and allowing the meta-atom to be designed in the upper left quadrant based on randomly selected parameters (such as the number, size, and position of the rectangular bars), the design is no longer limited to a fixed set of parameters, which greatly increases the diversity and complexity of the design and helps to achieve high efficiency and wide phase coverage. At the same time, mirror-copying the design in the upper left quadrant along the x-axis and y-axis to form the complete meta-atom design not only simplifies the design process but also ensures the consistency and symmetry of the design, which helps to improve the stability of the electromagnetic response.

[0110] At the same time, in the present invention, since each design can generate a unique meta-atom structure according to different random parameters, this diversified design method can effectively expand the working range of the model and make it no longer limited to a specific set of design parameters. Even in the face of changing design parameters, it can quickly adapt and generate corresponding designs, reducing the need to re-collect data and re-train the model.

[0111] Calculating the scattering parameters of the meta-atom design specifically includes:

[0112] Set the boundary conditions: Use the unit cell boundary conditions to simulate the electromagnetic response of the meta-atom design in the square lattice structure, and apply the open boundary conditions in the direction perpendicular to the plane of the meta-atom design, i.e., the z-direction, to ensure that the wavefront can propagate freely;

[0113] Define the illumination method: Use an x-polarized plane wave to irradiate the meta-atom design from the substrate side to simulate the incident light situation in the actual application scenario;

[0114] Determine the frequency range: Calculate the electromagnetic response of the meta-atom design in the frequency band of 30 to 60 THz, including transmittance and phase shift information;

[0115] Based on the set boundary conditions, irradiation mode, and frequency range, calculate the transmittance and phase shift of the meta-atom design through a full-wave electromagnetic simulation tool;

[0116] According to the obtained transmittance and phase shift data, calculate the phase and amplitude of the meta-atom design.

[0117] Train the initial network model through a data set to obtain a scattering parameter prediction model; predict the scattering parameters of the meta-atom design to be measured through the scattering parameter prediction model. Specifically, the initial network model is trained and tested by dividing the data set into a training set and a test set.

[0118] By constructing an initial network model including a splitting module, a first channel, a second channel, a stacking module, a comprehensive feature channel, and a calculation module, the present invention can effectively decompose the meta-atom design and extract its spatial features and physical features; at the same time, training the initial network model with meta-atom design samples of multiple different design structures and design parameters greatly improves the generalization ability of the model. The method of the present invention can accurately predict the real part and imaginary part of the transmission coefficient corresponding to the meta-atom design within a short time, and further calculate the scattering parameters, greatly improving the efficiency and accuracy of meta-atom scattering parameter prediction.

[0119] Specifically, the present embodiment trains the initial network model with the hyperparameters shown in Table 1 below:

[0120] Table 1:

[0121]

[0122] Embodiment 2

[0123] The scattering parameter prediction model proposed by the present invention is not limited to a single application scenario, and it can also be used as an optimization tool in the meta-surface or meta-atom design method based on a deep neural network (DNN). A significant advantage of the DNN-based design method is that it can generate a large number of design options at extremely low cost. However, all these designs need to go through a detailed characterization and evaluation process to determine the best solution, which is often more time-consuming than the generation of the design itself, especially when relying on complex simulation tools for performance verification.

[0124] The scattering parameter prediction model proposed by the present invention greatly alleviates the above problems with its fast response prediction ability. To demonstrate this application, as Figure 2 shown, the present embodiment of the invention also proposes a method for constructing a meta-device, and the construction method includes:

[0125] S1: Design a metasurface phase mask and decompose the performance requirements of the metasurface phase mask into those of each cell; one cell of the metasurface phase mask represents one meta-atom;

[0126] S2: Assign a serial number to the cell, with the initial value of the serial number equal to 1;

[0127] S3: Input the performance requirements of the cell corresponding to the current serial number into the pre-trained GAN model;

[0128] S4: Generate a meta-atom design set through the GAN model; the meta-atom design set includes multiple meta-atom designs;

[0129] S5: Input the meta-atom designs in the meta-atom design set into the scattering parameter prediction model as described above for prediction in sequence; based on the model prediction results corresponding to each meta-atom design in the meta-atom design set, obtain the optimal meta-atom design for the corresponding cell;

[0130] Step S5 specifically includes:

[0131] S50: Set the target number of iterations and initialize the number of iterations to 1;

[0132] S51: Input the meta-atom designs in the meta-atom design set into the scattering parameter prediction model as described above for prediction in sequence;

[0133] S52: Calculate the corresponding key performance indicators based on the model prediction results of each meta-atom design; the key performance indicators include transmission efficiency and phase shift accuracy.

[0134] S53: Select the best meta-atom design in the meta-atom design set as the current optimal design based on the key performance indicators; determine whether the current number of iterations is greater than or equal to the target number of iterations. If not, increment the number of iterations by 1 and proceed to the next step. If so, jump to step S55;

[0135] It should be noted that in order to select the best meta-atom design, this embodiment adopts an evaluation method based on key performance indicators. The key performance indicators mainly include transmission efficiency and phase shift accuracy. The transmission efficiency measures the transmission ability of the meta-atom to light waves, while the phase shift accuracy reflects the accuracy of the phase change generated by the meta-atom. In order to comprehensively evaluate the overall performance of each meta-atom design, this embodiment sets corresponding weights for these two indicators. By multiplying the transmission efficiency and phase shift accuracy by their respective weights and calculating the weighted sum, the comprehensive score of each meta-atom design can be calculated, and the design with the highest comprehensive score is selected as the current optimal design. This method can not only flexibly balance the importance of different performance indicators but also ensure that the finally selected design reaches the optimal or near-optimal performance in multiple dimensions.

[0136] S54: Feed the current optimal design back to the GAN model. The GAN model uses the current optimal design as a reference design, regenerates a new set of meta-atom designs, and returns to step S51;

[0137] S55: Set the current optimal design as the optimal meta-atom design and terminate the iteration.

[0138] In step S5 of the present invention, an iterative optimization mechanism is introduced. By continuously feeding the current optimal design back to the GAN model to generate a new set of meta-atom designs and repeating the evaluation and selection process until a predetermined number of iterations is reached. This method effectively ensures the superiority of the finally selected meta-atom design in key performance indicators, such as transmission efficiency and phase shift accuracy, etc., and realizes the maximization of the performance of the meta-device.

[0139] S6: Determine whether the current serial number is greater than or equal to the maximum serial number of the cells. If not, increment the serial number by 1 and return to step S3; if so, proceed to the next step;

[0140] S7: Assemble the meta-device through the optimal meta-atom designs of each cell.

[0141] The meta-device in this embodiment is specifically a transmissive meta-lens. Specifically:

[0142] This embodiment uses a trained generative adversarial network (GAN) model to design a transmissive meta-lens with an operating frequency of 57 THz. The meta-lens is composed of 50×50 meta-atoms, the lattice size is set to 2.6 μm, forming a device with a total area of 130 μm×130 μm. The focal length of this meta-lens is set to 80 μm, the numerical aperture (NA) is 0.63, and it is composed of meta-atoms made of a dielectric material with a refractive index of 4.7 and a thickness of 0.75 μm, placed on a substrate with a refractive index of 1.4. After calculating the required metasurface phase mask, a GAN model specifically for generating meta-atom designs is further trained to generate specific meta-atom designs for each cell within the transmissive meta-lens. Then, the generated meta-atom designs are evaluated using the scattering parameter prediction model proposed in the present invention.

[0143] This cascaded design-evaluation process undergoes multiple iterations to select the optimal meta-atom design for each cell and finally assemble it into a complete meta-device. To verify the prediction accuracy of the scattering parameter prediction model and the effectiveness of the entire design-evaluation process, as Figure 4As shown, in this embodiment, this combined network (i.e., the GAN model plus the scattering parameter prediction model) was used for 1, 2, 5, and 10 iterations, respectively generating four different transmissive metasurfaces, and their performances were tested using a full-wave simulation tool. The results are shown in Figure 4 where the transmissive metasurface is placed on the x-y plane and the optical axis is along the z direction. Figure 4 Figure 4 presents the two-dimensional electric field distributions simulated on the x-z plane of four different transmissive metasurfaces, as well as the one-dimensional electric field distribution along the x-axis on the focal plane. As the number of optimization iterations increases, the peak electric field intensity at the center of the focus gradually increases, indicating that during the optimization process, the scattering parameter prediction model identified the metaatom designs with higher transmission efficiency and more accurate phase shifts, thus demonstrating the effectiveness and superiority of the method of the present invention. It should also be specifically noted that Figure 4 in:

[0144] 1. E (V / m) represents the electric field strength (Electric Field), with the unit of volts per meter (V / m). In the left graph of the figure, the vertical axis indicates the electric field strength values, reflecting the electric field distribution at different positions.

[0145] 2. Z (μm) represents the position coordinate along the optical axis, with the unit of micrometers (μm). In the left graph of the figure, the horizontal axis indicates the position of the Z-axis, showing the variation of the electric field strength with the Z-axis.

[0146] 3. Number of iterations = 1, 2, 5, 10: corresponding to the four subgraphs in the figure respectively. Each subgraph shows the two-dimensional electric field distribution of the transmissive metasurface at different iteration numbers. As the number of iterations increases, the peak electric field amplitude at the center of the focus increases, indicating the effectiveness of the optimization process.

[0147] 4. X, Y, Z:

[0148] X and Y: represent the horizontal and vertical directions in the spatial coordinate system, used to describe the position within the plane where the transmissive metasurface is located.

[0149] Z: represents the optical axis direction, i.e., the direction of light propagation.

[0150] The X, Y, and Z axes are marked in the subgraph at the lower right corner to help understand the spatial orientation of the electric field distribution.

[0151] 5. Gradient color bar: represents the magnitude of the electric field strength, with the color changing from light to dark corresponding to the electric field strength increasing from low to high. The distribution of the electric field strength at different positions is visually shown through the color change.

[0152] 6. In the color bar at the upper left corner of the picture:

[0153] (1) Blue (labeled as 1): represents the electric field intensity distribution curve when the number of iterations is 1, that is, the change of the electric field intensity within the focal region;

[0154] (2) Orange (labeled as 2): represents the electric field intensity distribution curve when the number of iterations is 2;

[0155] (3) Yellow (labeled as 5): represents the electric field intensity distribution curve when the number of iterations is 5;

[0156] (4) Purple (labeled as 10): represents the electric field intensity distribution curve when the number of iterations is 10.

[0157] The present invention generates a variety of meta - atom designs by using a pre - trained GAN model, and uses the above - mentioned scattering parameter prediction model to evaluate the performance of each design, so as to select the optimal meta - atom design. This process ensures that each cell can meet specific performance requirements, thereby improving the design and manufacturing accuracy of the entire meta - device.

[0158] It should be noted that the term "including" and its variants used in the embodiments of the present invention are open - ended, that is, "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "multiple" mentioned in the embodiments of the present invention are illustrative rather than restrictive. Those skilled in the art should understand that unless clearly stated otherwise in the context, it should be understood as "one or more".

[0159] The term "embodiment" in this specification means that the specific features, structures or characteristics described in combination with the embodiment may be included in at least one embodiment of the present invention. The phrase appears in various positions in the specification does not necessarily mean the same embodiment, nor does it mean being independent or alternative to other embodiments and mutually exclusive. The various embodiments in this specification are described in a related manner, and the same or similar parts between the various embodiments are referred to each other. In particular, for the device, equipment, and system embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts refer to the partial description of the method embodiments.

[0160] The above - described embodiments only represent several implementation manners of the present invention. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of patent protection. It should be pointed out that for those of ordinary skill in the art, without departing from the concept of the present invention, several deformations and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention should be subject to the appended claims.

Claims

1. A method for predicting scattering parameters of meta-atoms, characterized in that Comprising: Construct an initial network model including a splitting module, a first channel, a second channel, a stacking module, a comprehensive feature channel, and a calculation module; where: The splitting module is used to decompose the meta-atom design into a two-dimensional image tensor and a one-dimensional attribute tensor; the meta-atom design includes a design structure and design parameters; the design structure corresponds to the two-dimensional image tensor, and the design parameters correspond to the one-dimensional attribute tensor; The first channel is used to extract the spatial features of the two-dimensional image tensor through convolution and pooling operations to obtain an image feature vector; The second channel is used to extract the physical features of the one-dimensional attribute tensor through non-linear transformation and spatial tiling operations to obtain an attribute feature vector; The stacking module is used to stack the image feature vector and the attribute feature vector to obtain a concatenated feature; The comprehensive feature channel is used to process the concatenated feature through convolution and pooling operations to obtain an electromagnetic response feature; The calculation module is used to obtain the real part and the imaginary part of the transmission coefficient corresponding to the meta-atom design according to the electromagnetic response feature, and calculate the scattering parameters based on the real part and the imaginary part of the transmission coefficient; where the scattering parameters include phase and amplitude; Generate multiple meta-atom designs with different design structures and design parameters, calculate the scattering parameters of the meta-atom designs, and label the corresponding meta-atom designs with the scattering parameters to obtain meta-atom design samples, and form a data set through the meta-atom design samples; the design parameters include: refractive index, thickness, and lattice size; The generating multiple meta-atom designs with different design structures and design parameters is specifically: Select a square canvas with a preset size as the design space for each meta-atom; For each meta-atom, design a preset number of rectangular bars in the upper left quadrant of the canvas based on randomly selected parameters; The randomly selected parameters include: the number, size, and position of the rectangular bars; each rectangular bar represents a set type of high refractive index material, and the rectangular bars correspond one-to-one to the types of high refractive index materials; Mirror-copy the design in the upper left quadrant along the x-axis and y-axis to form the complete design of each meta-atom; By adjusting the randomly selected parameters and repeating the above steps, generate multiple meta-atom designs with different design structures and design parameters; Train the initial network model through the data set to obtain a scattering parameter prediction model; predict the scattering parameters of the meta-atom design to be measured through the scattering parameter prediction model.

2. The scattering parameter prediction method for meta-atoms according to claim 1, wherein The first channel includes: A first convolutional layer, used to perform the first convolutional calculation on the two-dimensional image tensor to obtain a first spatial feature; A first max pooling layer, used to perform a max pooling operation on the first spatial feature to obtain a pooled feature; A second convolutional layer, used to perform the second convolutional calculation on the pooled feature to obtain a second spatial feature; A second max pooling layer, used to perform a max pooling operation on the second spatial feature to obtain an image feature vector.

3. A method for predicting scattering parameters of meta-atoms according to claim 2, characterized in that The second channel includes: A non-linear transformation layer, used to perform a non-linear transformation on the one-dimensional attribute tensor to obtain a transformed feature; A spatial tiling layer, used to perform a spatial tiling operation on the transformed feature to obtain an attribute feature vector.

4. A method for predicting scattering parameters for meta-atoms according to claim 3, characterized in that, The comprehensive feature channel includes, connected in sequence: The third convolutional layer is used to perform a convolutional operation on the concatenated features to obtain initial convolutional features; The third max pooling layer is used to perform a max pooling operation on the initial convolutional features to obtain initial pooling features; The fourth max pooling layer is used to perform a max pooling operation on the initial pooling features again to obtain target pooling features; The fourth convolutional layer is used to perform a convolutional calculation on the target pooling features and flatten the result obtained from the convolutional calculation to obtain target convolutional features; The first fully connected layer is used to perform a fully connected operation on the target convolutional features to obtain initial fully connected features; The second fully connected layer is used to transform the initial fully connected features to obtain electromagnetic response features.

5. A method for predicting scattering parameters for meta-atoms according to claim 4, characterized in that In the comprehensive feature channel, the third convolutional layer, the third max pooling layer, the fourth max pooling layer, the fourth convolutional layer, and the first fully connected layer all include batch normalization units and ReLU activation functions; The batch normalization unit is used to perform batch normalization on the features obtained after performing corresponding operations on their respective layers; The ReLU activation function is used to perform a non-linear transformation on the result of the batch normalization process to finally obtain the output features corresponding to that layer; the output features are initial pooling features, target pooling features, target convolutional features, or initial fully connected features.

6. A method for predicting scattering parameters for meta-atoms according to claim 1, characterized in that The scattering parameters designed by the computational meta-atom are specifically: Set boundary conditions: Use unit cell boundary conditions to simulate the electromagnetic response of the meta-atom design in a square lattice structure, and apply open boundary conditions in the direction perpendicular to the plane of the meta-atom design, i.e., the z-direction, to ensure that the wavefront can propagate freely; Define the illumination method: Use an x-polarized plane wave to illuminate the meta-atom design from the substrate side to simulate the incident light situation in an actual application scenario; Determine the frequency range: Calculate the electromagnetic response of the meta-atom design in the frequency band from 30 to 60 THz, including transmittance and phase shift information; Based on the set boundary conditions, illumination method, and frequency range, calculate the transmittance and phase shift of the meta-atom design through a full-wave electromagnetic simulation tool; According to the obtained transmittance and phase shift data, calculate the phase and amplitude of the meta-atom design.

7. A construction method of a meta-device, characterized in that, Include: S1: Design a metasurface phase mask and decompose the performance requirements of the metasurface phase mask into those of each cell; one cell of the metasurface phase mask represents one meta-atom; S2: Assign a serial number to the cell, and the initial value of the serial number is equal to 1; S3: Input the performance requirements of the cell corresponding to the current serial number into the pre-trained GAN model; S4: Generate a set of meta-atom designs through the GAN model; the set of meta-atom designs includes multiple meta-atom designs; S5: Input the meta-atom designs in the set of meta-atom designs into the scattering parameter prediction model in the method described in any one of claims 1 to 6 in sequence for prediction; According to the model prediction results corresponding to each meta-atom design in the set of meta-atom designs, obtain the optimal meta-atom design for the corresponding cell; S6: Determine whether the current serial number is greater than or equal to the maximum serial number of the cells. If not, increment the serial number by 1 and return to step S3; If so, proceed to the next step; S7: Assemble a meta-device through the optimal meta-atom design of each cell.

8. A method for constructing a meta-device according to claim 7, characterized in that, The S5 step specifically includes: S50: Set the target number of iterations and initialize the number of iterations to 1; S51: Input the meta-atom designs in the meta-atom design set into the scattering parameter prediction model in the method described in any one of claims 1 to 6 in sequence for prediction; S52: Calculate the corresponding key performance indicators based on the model prediction results of each meta-atom design; S53: Select the best meta-atom design in the meta-atom design set as the current optimal design based on the key performance indicators; Determine whether the current number of iterations is greater than or equal to the target number of iterations. If not, increment the number of iterations by 1 and proceed to the next step. If so, jump to step S55; S54: Feed the current optimal design back to the GAN model. The GAN model uses the current optimal design as a reference design to regenerate a new meta-atom design set and return to step S51; S55: Then set the current optimal design as the optimal meta-atom design and terminate the iteration.

9. A method for constructing a meta-device according to claim 8, characterized in that, The key performance indicators include transmission efficiency and phase shift accuracy.

Images