Resonance metasurface design method and device based on spectrum sensing improvement

Abstract

In order to solve the problems of long design time consumption, low efficiency, high calculation cost and insufficient prediction precision caused by information loss in the prior art, the invention provides an improved resonant metasurface design method and device based on spectrum sensing. The method comprises the following steps: designing an improved GLSAT forward prediction network based on spectrum sensing; the GLSAT forward prediction network is trained and optimized; a DNN reverse design network is designed; the DNN reverse design network and a trained GLSAT forward prediction network are cascaded, and a cascaded reverse design network is obtained; preprocessing the ideal spectrum by using GSSG, inputting the preprocessed spectrum into a cascade reverse design network, and training and optimizing the DNN; and completing the design of the resonance metasurface by using the optimized DNN reverse design network. The method provided by the invention has the characteristics of short time consumption, high efficiency and low calculation cost while improving the design prediction precision.

Description

Technical Field

[0001] The present invention relates to the field of resonant metasurface design, in particular to the field of resonant metasurface inverse design based on spectral perception characteristics, and specifically to a resonant metasurface design method and device based on spectral perception improvement. Background Art

[0002] In recent years, all-dielectric resonant metasurfaces have attracted much attention in the design of optical components due to their advantages of low loss and high quality factor (Q factor). Such resonant metasurfaces are widely used in reflection, transmission, absorption filters and phase control devices. However, studying the spectral characteristics of resonant metasurfaces faces multiple challenges: on the one hand, their spectral response usually exhibits complex behavior, involving multiple resonant modes and nonlinear effects, which makes prediction and design complicated; on the other hand, the continuous change of spectral characteristics over a wide frequency range produces a high-dimensional data set, and effective analysis and optimization of it requires advanced computing technology.

[0003] Traditional resonant metasurface design relies on numerical full-wave simulation (such as finite element method FEM, finite difference time domain method FDTD and finite integration method FIT). Although this on-demand resonant metasurface design method has high accuracy, it still relies on empirical manual parameter adjustment and trial and error, which requires a lot of computing resources and time. Especially in the case of complex geometric structures, it is very time-consuming to achieve the design of specific electromagnetic response through empirical parameter adjustment and repeated optimization, resulting in low efficiency. In addition, with the rapid development of artificial intelligence (AI), especially machine learning (ML), these technologies have provided powerful computational tools to overcome the above obstacles and promoted the related research on the inverse design of resonant metasurfaces. Evolutionary algorithms such as particle swarm optimization (PSO) and genetic algorithm (GA) are often used for resonant metasurface design. However, these heuristic algorithms are often time-consuming and computationally inefficient when involving a large number of parameters due to the complex calculation process, making it difficult to meet real-time design requirements.

[0004] In contrast, deep learning models based on deep neural networks (DNNs) are able to solve complex problems with the powerful ability of multiple layers of hidden units. This ability allows them to reveal the hidden relationship between structural parameters and their electromagnetic (EM) responses (such as amplitude and phase). Therefore, a large database generated by full-wave simulation can be used to train neural networks. After the model is trained with high accuracy, the data-driven approach can shorten the forward prediction time from structural parameters to target amplitude or phase to milliseconds, thereby accelerating the reverse design process from target spectrum to structural design.

[0005] Nevertheless, the inverse prediction network from spectrum to structure may produce non-unique solutions because the same spectral input may correspond to multiple structural outputs. In addition, the target spectrum is usually idealized (ideal spectrum), which makes it difficult for the model to accurately capture the underlying complex relationship. Some existing technologies use generative adversarial networks (GANs) to find corresponding structures from the target spectrum. On the other hand, networks such as fully connected networks (FCNs), bidirectional neural networks, and tensor neural networks (NTNs) have also been applied to the inverse design of spectral prediction of resonant metasurfaces. Some existing technologies use dimensionality reduction methods to solve the matching problem between structural geometry and spectral data. Although unnecessary sharp peaks can be removed, it also leads to the loss of spectral information.

[0006] Transformer models based on self-attention mechanisms have shown excellent performance in language and vision tasks in recent years. The emergence of this architecture overcomes many limitations of previous models and can successfully capture key spectral features from low-dimensional structures, including high-Q-factor resonant metasurface sensor resonances, broadband solar metamaterial absorbers, and molecular fingerprint features in Raman spectroscopy. However, the application of the Transformer architecture usually requires a large number of stacked layers to achieve the expected performance, which leads to increased model complexity and computational cost. These challenges indicate that more efficient Transformer designs are needed to reduce the number of layers while maintaining or improving the accuracy of predicting electromagnetic responses. Summary of the invention

[0007] In view of this, in order to solve the problems of long design time, low efficiency, high calculation cost, and insufficient prediction accuracy caused by information loss in the prior art, the present invention provides a resonant metasurface design method and device based on spectral perception improvement. The resonant metasurface design method based on spectral perception improvement has the characteristics of short time consumption, high efficiency and low calculation cost while improving the prediction accuracy of the algorithm.

[0008] The present invention provides a resonant metasurface design method based on spectral perception improvement, comprising the following steps: Step 110: Design a GLSAT forward prediction network to achieve forward prediction from resonant metasurface structural parameters to ideal spectra to obtain predicted spectra; the design of the GLSAT forward prediction network is based on a multi-head attention mechanism improved by spectral perception; the GLSAT forward prediction network includes an FCL module for dimensional expansion and a dual Transformer module for feature decomposition and extraction of high-dimensional spectra; Step 120: According to the prediction design requirements for achieving an ideal spectral response, a data set of resonant metasurface structure parameters is selected, and the GLSAT forward prediction network is trained and optimized to obtain a trained GLSAT forward prediction network; Step 130: Design a DNN reverse design network to achieve the design of resonant metasurface structure parameters that meet the ideal spectral response; the DNN reverse design network implements the design algorithm based on a deep neural network; Step 140: Cascading the DNN reverse design network with the trained GLSAT forward prediction network to construct a cascade reverse design network from an ideal spectrum to resonant metasurface structural parameters; the cascade reverse design network includes a Gaussian line spectrum generator for preprocessing the ideal spectrum, a DNN reverse design network, and a trained GLSAT forward prediction network; Step 150: inputting the ideal spectrum as a training set into the cascade inverse design network, training and optimizing the DNN inverse design network, until a cascade inverse design network that meets the design accuracy requirements is obtained; Step 160: Input a set of ideal spectra based on actual design requirements into the cascaded inverse design network that meets the design accuracy requirements, and use the DNN inverse design network in the cascaded inverse design network to generate resonant metasurface structural parameters that meet the ideal spectral response, thereby completing the design of the resonant metasurface.

[0009] Specifically, the FCL module includes multiple residual modules with a fully connected layer structure, which are used to achieve three-stage dimensional expansion. The fully connected layer structure includes multiple linear layers; the residual modules include at least three types: a first residual module, a second residual module and a third residual module; each linear layer of any residual module is followed by an activation function LeakyReLU, which is used to introduce nonlinear factors to alleviate gradient disappearance, and its negative slope is ; The dual Transformer module is used to extract global and local features of high-dimensional spectral data, and extract all features through a multi-head attention mechanism; the dual Transformer module includes a GPT module for global perception and an LPT module for local perception; the multi-head attention mechanism includes a cross-attention mechanism for the GPT module to extract global features to realize interactive multi-head attention analysis of spectral fragments, and a self-attention mechanism for the LPT module to extract local features to realize inline multi-head attention analysis of spectral fragments.

[0010] Furthermore, in Step 110, the GLSAT forward prediction network is used to achieve forward prediction from the resonant metasurface structure parameters to the ideal spectrum, and the process of obtaining the predicted spectrum includes: Step 111: Dimensionally expand the resonant metasurface structure parameters using the FCL module, including: The parameters of the dimensional resonant metasurface structure are expanded to The two first residual modules of represents the number of resonant metasurface structure parameters; the first residual module is expanded to The second residual module is further expanded to dimensional third residual module, where represents the spectral dimension; Step 112: preprocessing the output data of the FCL module before entering the dual Transformer module, including: using a layer of dimension The fully connected layer structure is used to standardize the data. is the standardized spectral dimension; the standardized data is processed through the projection matrix to obtain the query vector required by the multi-head attention mechanism , key vector Sum value vector ,in , is the total number of tokens in the spectrum, and token represents a spectrum fragment; Step 113: input the data preprocessed by Step 112 into the dual Transformer module for data processing, including: data processing by the GPT module, data processing by the LPT module, and smoothing the data by a one-dimensional convolution layer connecting the GPT module and the LPT module, and the LPT module and the data output layer respectively; Step 114: The data smoothed by the one-dimensional convolution layer connecting the LPT module and the data output layer is output through the output layer to obtain a predicted spectrum.

[0011] Preferably, in Step 113, The data processing process of the GPT module includes: Divide the data into Tokens; Use cross attention mechanism to capture the attention of token interaction, attention score and attention results Calculated by the following formulas:

[0012]

[0013] in, Representation Matrix The transpose of

[0014]

[0015]

[0016] Indicates the transpose of a matrix, and the Softmax function is used to convert a matrix data into a probability value score; By focusing all the heads Splicing together to obtain the multi-head attention analysis result based on the cross-attention mechanism is given by the following formula:

[0017] Among them, the Contact function is a tensor concatenation function used to achieve multiple attention results splicing; It is The attention result of heads, the number of heads is , is the weight parameter matrix associated with all heads; The obtained multi-head attention analysis results , passed to a normalized feed-forward layer to implement multi-layer normalization processing of data; The data processing process of the LPT module includes: Use Contains The self-attention mechanism of each attention head will perform multi-head attention analysis on a single token to obtain an attention score , given by the following formula:

[0018]

[0019] in, ,all and The data are all from the same spectral fragment. The Sigmoid function is an S-shaped activation function, and its output value is between; use and The Hadamard product of the attention result is obtained , is given by:

[0020] Among them, all , and The data are all derived from the same spectral fragment; Use the Contact function to collect the attention results of all heads Putting them together, we get the multi-head attention analysis result based on the self-attention mechanism, which is given by the following formula:

[0021] in, Indicates the number of heads, It is The attention result of each head, is the weight parameter matrix associated with all heads; The obtained multi-head attention analysis results , passed to a normalized feed-forward layer to implement multi-layer normalization processing of data.

[0022] Preferably, the normalized feedforward layer includes two normalization layers, a feedforward network, and an activation function LeakyReLU; the two normalization layers are respectively a first normalization layer and a second normalization layer; the structure of the normalized feedforward layer from top to bottom is: a first normalization layer, used to ensure the stability of feature distribution and improve the convergence performance of the model; a feedforward network, used to perform nonlinear transformation and further extract features; a second normalization layer, used to perform stable gradient propagation; an activation function LeakyReLU, used to introduce nonlinear factors to prevent gradient disappearance; During the multi-layer normalization processing of the data, after the data is transmitted from the first normalization layer to the feedforward network, the input and output of the feedforward network are fused by a residual connection to ensure that the input information of the data is retained.

[0023] Specifically, in Step 120, according to the prediction design requirements for achieving an ideal spectral response, a data set of resonant metasurface structure parameters is selected, and the process of training and optimizing the GLSAT forward prediction network includes: Step 121: Divide the data set into a training set, a validation set, and a test set in proportion; Step 122: The GLSAT forward prediction network is trained and optimized using the AdamW optimizer. The initial value of the learning rate in the AdamW optimizer is set to 0.0005. The learning rate is adjusted by cosine annealing scheduling and gradually decreases according to the cosine function. An L2 regularization function module is added to the AdamW optimizer to reduce overfitting. The loss function in the AdamW optimizer is The mean square error is used to measure the predicted spectrum. Compared with the original spectrum obtained by simulation The difference between the loss function Given by: , in, Indicates the number of samples in the training set; Step 123: Through the indicators on the validation set , evaluate the optimization effect of the GLSAT forward prediction network; the index The mean absolute error is calculated to provide a robust evaluation result of network optimization performance. Given by: , in, Indicates the number of samples in the validation set.

[0024] Specifically, in Step 130, the input layer of the DNN reverse design network includes spectral data, which is spectral points, the output layer generates a vector, storing the corresponding data of the resonant supersurface structure parameters; the deep neural network is composed of three consecutive fully connected hidden layers with a fully connected layer structure, and the three fully connected hidden layers respectively contain , and A neuron.

[0025] Specifically, in Step 150, the ideal spectrum as a training set is input into the cascade inverse design network, and the DNN inverse design network is trained and optimized until a cascade inverse design network that meets the design accuracy requirements is obtained, including: Step 151: preprocessing the ideal spectrum using the Gaussian line spectrum generator to obtain a Gaussian spectrum that is closer to a physically achievable spectrum, including constructing a set of Gaussian-like spectrum data sets to obtain a Gaussian spectrum corresponding to the response of the ideal spectrum; The ideal spectrum is obtained by filtering the original spectrum data set through a predefined requirement threshold to generate an ideal spectrum containing only 0 or 1 binarization; the screening threshold is set to 0.9; Step 152: Input the Gaussian spectrum into the cascade architecture of the DNN reverse design network and the trained GLSAT forward prediction network, and train and optimize the DNN reverse design network, including: inputting the Gaussian spectrum into the DNN reverse design network, using the DNN reverse design network to design, generate and output resonant supersurface structure parameters; using the output resonant supersurface structure parameters as input data of the trained GLSAT forward prediction network to generate and output a predicted spectrum; comparing the input Gaussian spectrum with the output predicted spectrum, calculating the loss function, and optimizing the parameters of the deep neural network by minimizing the gradient descent of the loss function, thereby improving the accuracy of the DNN reverse design network in designing resonant supersurface structure parameters.

[0026] Preferably, the predicted design requirements of the ideal spectrum meet the requirements of the designed resonant metasurface on the wavelength and intensity of the reflection peak, including a narrow-band filtering reflection resonant metasurface specially designed for an alkali metal laser with a line width of 10 nm and a wavelength of 795 nm; The data set selected in Step 120 is the resonant supersurface structure parameters of the square lattice diamond structure silicon dioxide substrate, including: the diameter of the cylinder ,high The distance between adjacent cylinders , where the diameter Less than distance , the dimension of the resonant metasurface structure parameters is , set the training rounds ; In the Step 121, the data set is divided into a training set, a validation set, and a test set according to a ratio of 80%, 10%, and 10%; In Step 151, by defining the standard deviation , parameterize the Gaussian line-shaped reflectance spectrum curve and construct a set of Gaussian-like spectrum data sets as Gaussian spectrum; represents the Gaussian spectrum obtained by the Gaussian line shape processor; The process of training the DNN reverse design network in Step 152 includes: Iteratively updating the weights and biases of each fully connected hidden layer in the deep neural network by minimizing the Euclidean distance between the input Gaussian spectrum and the predicted spectrum generated by the trained GLSAT forward prediction network, while keeping the network parameters of the trained GLSAT forward prediction network unchanged, and gradually optimizing the network parameters in the deep neural network until an optimized DNN reverse design network is obtained; Introducing weight factors WS , to enhance the impact of peak features on prediction results, so that the network can preferentially capture key spectral characteristics. W S Given by: ; The DNN inverse design network is trained using a loss function defined by the following formula : , in, Represents the number of training set samples of the DNN reverse design network, represents the resonant metasurface structure parameters generated by the DNN inverse design network, = represents the resonant metasurface structure parameter value selected in Step 120; The optimization performance index of the DNN reverse design network is defined by the following formula: ; In Step 160, the optimized DNN reverse design network is used to generate the resonant metasurface structure parameters with high reflectivity in the target band at one time according to the predicted design requirements of the ideal spectrum.

[0027] In addition, the present invention also protects a resonant metasurface design device based on spectral perception improvement, the device uses the steps of the above method to design a resonant metasurface, and the device includes the following modules: The first module is used to design a GLSAT forward prediction network to achieve forward prediction from the resonant metasurface structure parameters to the ideal spectrum and obtain the predicted spectrum; the design of the GLSAT forward prediction network is based on the multi-head attention mechanism improved by spectral perception; the GLSAT forward prediction network includes an FCL module for dimensional expansion and a dual Transformer module for feature decomposition and extraction of high-dimensional spectra; The second module is used to select a data set of resonant metasurface structure parameters according to the predicted design requirements for achieving an ideal spectral response, train and optimize the GLSAT forward prediction network, and obtain a trained GLSAT forward prediction network; The third module is used to design a DNN reverse design network to achieve the design of resonant metasurface structure parameters that meet the ideal spectral response; the DNN reverse design network implements the design algorithm based on a deep neural network; The fourth module is used to cascade the DNN reverse design network with the trained GLSAT forward prediction network to construct a cascade reverse design network from the ideal spectrum to the resonant metasurface structure parameters; the cascade reverse design network includes a Gaussian line spectrum generator for preprocessing the ideal spectrum, a DNN reverse design network and a trained GLSAT forward prediction network; The fifth module is used to input the ideal spectrum as a training set into the cascade inverse design network, train and optimize the DNN inverse design network, until a cascade inverse design network that meets the design accuracy requirements is obtained; The sixth module is used to input a set of ideal spectra based on actual design requirements into the cascaded inverse design network that meets the design accuracy requirements, and use the DNN inverse design network in the cascaded inverse design network to generate resonant metasurface structural parameters that meet the ideal spectral response, thereby completing the design of the resonant metasurface.

[0028] In summary, the present invention provides a resonant metasurface design method and device based on spectral perception improvement. Compared with the prior art, the method of the present invention mainly achieves the following improvements and significant effects: 1) By designing a forward prediction network with a GLSAT forward prediction network, and testing and optimizing the GLSAT forward prediction network, the prediction process from the resonant metasurface structure parameters to the ideal spectrum is accelerated, especially the dual Transformer module design in the GLSAT forward prediction network: using the GPT module therein, a cross-attention mechanism is used to perform feature extraction and multi-head attention analysis on the global perception of the spectrum; using the LPT module therein, a self-attention mechanism is used to perform feature extraction and multi-head attention analysis on the local perception of the spectrum. The dual Transformer module design achieves significant performance improvement in quickly capturing features such as global correlation and local details of high-dimensional spectral data.

[0029] 2) By considering the different requirements for high reflectivity of single-band reflection peaks and dual-band reflection peaks in the ideal spectrum, a DNN reverse design network that can handle different targets is designed to achieve reverse design optimization; by inputting the prediction results of the GLSAT forward prediction network and participating in the testing and optimization of the DNN reverse design network, the DNN reverse design network is cascaded with the optimized GLSAT forward prediction network, thereby further improving the prediction accuracy of spectral perception.

[0030] Therefore, the algorithm complexity of the aforementioned method of the present invention is limited. On the basis of saving computing costs, the efficiency and accuracy of the resonant metasurface design based on spectral perception improvement are improved, and it is expected to be applied to the development of fast and precise resonant metasurface devices. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A flowchart of a resonant metasurface design method based on spectral perception improvement provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the algorithm architecture of the GLSAT forward prediction network in the first embodiment of the present invention, wherein: and They are all data dimensions. is the dimension of the normalized spectral data, is the distance between the periodic arrangement of the resonant metasurface units. The global perception Transformer is the GPT module for global perception, and the local perception Transformer is the LPT module for local perception. Represents the attention score obtained in the GPT module, is the number of headers in the GPT module, Represents the attention score obtained in the LPT module, is the number of headers in the LPT module; Figure 3 Schematic diagram of the algorithm architecture of the dual Transformer module in the first embodiment of the present invention, wherein ProjectMatrix represents the projection matrix, , is the key vector, is a value vector, in the GPT module , in the LPT module , and Represent the attention results in the GPT module and LPT module respectively, W G is the weight parameter matrix related to all heads, Intra-block represents the cross-interaction between data modules, Intro-block represents the internal interaction of data modules, Contact is the splicing function, Coarse-tune is the coarse tuning layer, Fine-tune is the fine tuning layer, and the data dimensions of the coarse tuning layer and the fine tuning layer are , Layer Norm 1 is the first normalization layer, Layer Norm 2 is the second normalization layer, Feed Forword is a feedforward network, and LeakyReLU is an activation function that introduces nonlinearity; Figure 4 Schematic diagram of the algorithm architecture of the DNN reverse design network in the first embodiment of the present invention, wherein GSSG is a Gaussian Linear Spectrum Generator; Figure 5Schematic diagram of the algorithm architecture of the cascaded inverse design network in the first embodiment of the present invention, wherein Loss represents the error between input and output data characterized by the loss function, and GLSAT refers to the GLSAT forward prediction network; Figure 6 Schematic diagram of the algorithm architecture of the DNN reverse design network in the second embodiment of the present invention, wherein: is the diameter of the cylinder, is the height and is The distance between adjacent cylinders; Figure 7 A schematic diagram of the algorithm architecture of a DNN reverse design network in the second embodiment of the present invention; Figure 8 Schematic diagram of the comparison between the original spectrum and the predicted spectrum obtained by simulation for the second embodiment of the present invention, wherein MAE is the mean absolute error, (a)-(h) respectively represent the comparison curves of the original spectrum and the predicted spectrum obtained by simulation when MAE is 0.0032, 0.0067, 0.0042, 0.0034, 0.0048, 0.055, 0.0020 and 0.0031, is the original spectrum, is the predicted spectrum; Fig. 9 The absolute value error range distribution histogram of the spectral points obtained by simulation of the second embodiment of the present invention; Fig.10 This is a schematic diagram comparing the data of the Gaussian spectrum obtained by simulating different structural parameter conditions in the second embodiment of the present invention with the results of the predicted spectrum, wherein: represents the data of a Gaussian spectrum, represents the predicted spectrum, is the original spectrum, is a simulated spectrum. DETAILED DESCRIPTION

[0032] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. For example, the words "upper" and "lower" used to describe the orientation are only used to describe the network algorithm architecture involved in the method steps of the present invention with respect to the corresponding positional relationship in the accompanying drawings of the present invention, and should not be used to limit the technical features involved in the method or device of the present invention.

[0033] In the first embodiment, referring to Figure 1 As shown, the present invention proposes a resonant metasurface design method based on spectral perception improvement, and the method specifically comprises the following steps: Step 110: Design a GLSAT forward prediction network to achieve forward prediction from the resonant metasurface structure parameters to the ideal spectrum to obtain the predicted spectrum; the design of the GLSAT forward prediction network is based on a multi-head attention mechanism improved by spectral perception; the GLSAT forward prediction network includes an FCL module for dimensional expansion and a dual Transformer module for feature decomposition and extraction of high-dimensional spectra; Step 120: According to the prediction design requirements for achieving an ideal spectral response, a data set of resonant metasurface structure parameters is selected, and the GLSAT forward prediction network is trained and optimized to obtain a trained GLSAT forward prediction network; Step 130: Design a DNN reverse design network to achieve the design of resonant metasurface structure parameters that meet the ideal spectral response; the DNN reverse design network implements the design algorithm based on a deep neural network; Step 140: Cascading the DNN reverse design network with the trained GLSAT forward prediction network to construct a cascade reverse design network from an ideal spectrum to resonant metasurface structural parameters; the cascade reverse design network includes a Gaussian line spectrum generator for preprocessing the ideal spectrum, a DNN reverse design network, and a trained GLSAT forward prediction network; Step 150: inputting the ideal spectrum as a training set into the cascade inverse design network, training and optimizing the DNN inverse design network, until a cascade inverse design network that meets the design accuracy requirements is obtained; Step 160: Input a set of ideal spectra based on actual design requirements into the cascaded inverse design network that meets the design accuracy requirements, and use the DNN inverse design network in the cascaded inverse design network to generate resonant metasurface structural parameters that meet the ideal spectral response, thereby completing the design of the resonant metasurface.

[0034] Specifically, in Step 110, the algorithm architecture of the GLSAT forward prediction network is as follows: Figure 2 As shown, the FCL module uses multiple residual modules with a fully connected layer structure to achieve three-stage dimensional expansion, wherein the fully connected layer structure includes multiple linear layers; the residual modules include at least three types: a first residual module, a second residual module and a third residual module; the three-stage dimensional expansion specifically includes: first, the initial Dimensional resonant metasurface structure parameters are extended to dimensions The two first residual modules of Represents the number of structural parameters of the resonant metasurface; then the first residual module is expanded to a dimension The second residual module is further expanded to a dimension The third residual module is Represents the spectral dimension. Each linear layer is followed by an activation function LeakyReLU that introduces nonlinear factors, and its negative slope is .

[0035] Furthermore, the dual Transformer module includes a global perception Transformer (hereinafter referred to as the GPT module) and a local perception Transformer (hereinafter referred to as the LPT module). The GPT module is used to extract the correlation between spectral segments, and the LPT module is used to refine the spectral details within each spectral segment. The design of the dual Transformer module decomposes the high-dimensional spectral data into global features and local features, and extracts all features through a multi-head attention mechanism. Specifically, the global features are extracted by performing interactive multi-head attention analysis of spectral segments through the cross-attention mechanism of the GPT module, and the local features are extracted by performing inline multi-head attention analysis of spectral segments through the self-attention mechanism of the LPT module, thereby achieving significant performance improvement in quickly capturing features such as global correlation and local details of high-dimensional spectral data.

[0036] like Figure 2 As shown in FIG. 1 , before the output data of the FCL module enters the dual Transformer module, it first passes through a fully connected layer structure for data standardization, and its data dimension is , and then the data is processed by the projection matrix (ProjectMatrix) to obtain the query vector required by the multi-head attention mechanism , key vector Sum value vector , ,in is the total number of spectral fragments (tokens).

[0037] Further, if Figure 2 As shown, the output data of the FCL module is processed by normalization and projection matrix and then enters the dual Transformer module for processing. This process includes: (i) The first stage, such as Figure 3 As shown, the data enters the GPT module for processing, including dividing the spectral data into tokens, and then perform multi-head attention analysis on these tokens through the cross-attention mechanism.

[0038] Here, the cross attention mechanism is used to capture the attention of token interaction, and the attention score and attention results Calculated by the following formulas (1) and (2): (1) (2) in, Representation Matrix The transpose of (3) here represents the transpose of the matrix; the attention score By calculating the matrix and Transpose the dot product and divide by The Softmax function is a commonly used function in machine learning and deep learning, and is used to convert a vector or matrix data into a probability value.

[0039] When performing matrix multiplication, all Will be with Data interaction occurs, and the result of multi-head attention analysis is obtained by combining the attention results of all heads Spliced ​​together, it is calculated according to the following formula (4): (4) Among them, the Contact function is a tensor concatenation function used to achieve multiple attention results splicing; It is The attention result of heads, the number of heads is , is a weight parameter matrix associated with all heads. The multi-head based cross attention mechanism enables the GLSAT forward prediction network to quickly capture the global dependencies between spectral segments, thereby providing coarse tuning for global spectral perception.

[0040] (ii) The second stage, such as Figure 3 The data shown is processed by the LPT module, including the use of m The self-attention mechanism of each attention head performs multi-head attention analysis on a single token. This self-attention mechanism can quickly capture the internal relationship of each spectral fragment.

[0041] The calculation of the attention score is similar to that of the GPT module, but the data used in the query vector, key vector, and value vector are all derived from the same spectral fragment. The calculation result is obtained through the Sigmoid function, as shown in the following formulas (5) and (6): (5) (6) The Sigmoid function is an S-shaped activation function, and its output value is between and

[0042] Attention result is described by the Hadamard product of and : (7) In the above calculation process, only contains the information of and , while contains the information of , and , where , , and all come from the same spectral segment.

[0043] Subsequently, the attention results of all heads are concatenated using the Contact function to obtain the result of multi-head attention analysis, as shown in the following formula (8): (8) Among them, represents the number of heads, is the attention result of the -th head, and is the weight parameter matrix related to all heads. The multi-head based self-attention mechanism quickly captures the internal connections within each spectral segment, thereby achieving fine adjustment of the spectrum for local spectral perception.

[0044] Further, as Figure 3As shown, the data processing process of the GPT module or LPT module includes: after obtaining the multi-head attention analysis results, the results are adjusted by the coarse-tune layer (Coarse-tune) or fine-tune layer (Fine-tune) in the system and then passed to a normalized feedforward (Norm and Feed Forward, NFF) layer to achieve multi-layer normalization of data. The normalized feedforward layer has a multi-layer structure, which specifically includes from top to bottom: the first normalization layer (Layer Norm 1), which is used to ensure the stability of feature distribution and improve the convergence performance of the model; a feedforword network (FeedForword Network, FFN) composed of fully connected layers, which is used to perform nonlinear transformations and further extract features; the second normalization layer (LayerNorm 2), which is used to perform stable gradient propagation; an activation function LeakyReLU (Leaky Rectified LinearUnit), which is used to introduce nonlinearity and prevent gradient disappearance. When the data enters the NFF for multi-layer normalization processing, after the data is transmitted from the Layer Norm 1 to the FFN, the input and output of the FFN are fused by residual connection to ensure that the input information of the data can be retained.

[0045] (iii) Finally, if Figure 2 As shown, between the GPT module and the LPT module, as well as between the LPT module and the data output layer, a one-dimensional convolution layer is used to smooth the data results output by the NFF to reduce high-frequency fluctuations in the predicted spectral curve.

[0046] Preferably, the process of training and optimizing the GLSAT forward prediction network in Step 120 includes: Step 121: According to the design requirements for achieving an ideal spectral response, a data set of resonant metasurface structure parameters is selected, and the data set is divided into a training set, a validation set, and a test set in proportion; Step 122: The GLSAT forward prediction network is trained and optimized using the AdamW optimizer, and the initial value of the learning rate in the AdamW optimizer is set to 0.0005; an L2 regularization (weight decay) function module is added to the optimizer to reduce overfitting; Step 123: Evaluate the optimization effect of the GLSAT forward prediction network.

[0047] Specifically, during the training of the GLSAT forward prediction network, the learning rate is adjusted by cosine annealing scheduling and gradually decreases according to the cosine function. The period of the cosine annealing scheduling is consistent with the total number of training rounds, so that the learning rate is The weight decay is gradually decayed in each training round. This strategy can help the model converge smoothly by making large updates at the beginning of training and then gradually refining the weights to approach the optimal solution. The use of the AdamW optimizer can improve the convergence speed and generalization ability by decoupling weight decay from gradient updates, so that weight decay is applied more effectively in each iteration. The loss function in the AdamW optimizer The mean square error (MSE) is used to measure the predicted spectrum The original spectrum obtained from the actual simulation The difference between the loss function Given by: , (9) in, Indicates the number of samples in the training set.

[0048] Furthermore, when evaluating the optimization effect of the GLSAT forward prediction network, the mean absolute error (MAE) is used to calculate the index on the validation set. , used to provide a robust evaluation result of network optimization performance, the indicators Given by: (10) in, Indicates the number of samples in the validation set.

[0049] Specifically, the algorithm architecture of the DNN reverse design network described in Step 130 is as follows: Figure 4 As shown, the DNN reverse design network is based on a deep neural network (DNN) and is used to design a resonant metasurface structure that can achieve an ideal spectral response. The input layer of the DNN reverse design network contains spectral data. The output layer generates a reflection spectrum point. Vector, storing the corresponding data of the resonant supersurface structure parameters. The DNN consists of three consecutive fully connected hidden layers (linear layers), and the three linear layers contain , and A neuron.

[0050] Specifically, in Step 140, in order to solve the non-convergence problem caused by non-unique solutions, the fully trained GLSAT forward prediction network is cascaded with the DNN inverse design model and used together with the Gaussian-shape spectrum generator (GSSG), as shown in Figure 5 As shown, a cascaded reverse design network is constructed. Figure 5The algorithmic architecture of the cascaded inverse design network is presented, in which GSSG is used to establish the correspondence between the response of an ideal spectrum and a spectrum close to that which is physically achievable.

[0051] In order to obtain an ideal spectrum, it is necessary to I demand ) The original spectral data set is filtered to generate an ideal spectrum containing only 0 or 1 binarization; the filtering threshold can be set to 0.9.

[0052] Specifically, in Step 150, the ideal spectrum as a training set is input into the cascade inverse design network, and the DNN inverse design network is trained and optimized until a cascade inverse design network that meets the design accuracy requirements is obtained, including: Step 151: Preprocess the ideal spectrum using the GSSG to obtain a Gaussian spectrum: by constructing a set of Gaussian-like spectral data sets that correspond to the responses of the ideal spectrum, the gap between the ideal spectrum and the physically achievable spectrum is reduced.

[0053] Step 152: Input the Gaussian spectrum Figure 5 The cascade architecture of the DNN reverse design network and the trained GLSAT forward prediction network is shown, and the DNN therein is trained and optimized, including: Inputting the Gaussian spectrum into a DNN reverse design network, using the DNN reverse design network to design and generate resonant metasurface structure parameters, and outputting data; Using the above output data as input data of the trained GLSAT forward prediction network to generate a predicted spectrum; The input Gaussian spectrum is compared with the output predicted spectrum to calculate the loss function, and the parameters of the DNN are optimized by minimizing the gradient descent of the loss function, thereby improving the accuracy of the parameter design of the resonant metasurface structure in the DNN inverse design network.

[0054] In the second embodiment of the present invention, a narrowband filtering reflection resonant metasurface for an alkali metal laser with a line width of 10 nm and a wavelength of 795 nm is specially designed through the steps of the resonant metasurface design method described in the first embodiment. In other words, the prediction demand for the ideal spectral response is mainly to meet the requirements of the designed resonant metasurface on the wavelength and intensity of the reflection peak. At this time, the ideal spectrum reflects the characteristic requirements of the reflection spectrum. Therefore, the data set selected in Step 120 is the structural parameters corresponding to the resonant metasurface of the square lattice diamond structure silicon dioxide substrate, and the structural parameters include: the diameter of the cylinder ,high The distance between adjacent cylinders The resonant metasurface is composed of a plurality of resonant metasurface units with the same structure arranged periodically. The structural parameters are as follows: Figure 6 Medium Coordinate System As shown in the structure diagram of the resonant metasurface unit, the dimension variables of the structural parameters are 3. Dimension variables , set the training rounds . Figure 7 The algorithm architecture of the cascade reverse design network constructed in this embodiment. Considering the actual manufacturing limitations, and the diameter Must be smaller than the distance of the cylinder periodic arrangement , a systematic parameter scan is performed on the three independent structural parameters and uniformly sampled within a defined range. The scan range is shown in Table 1.

[0055] Table 1 Structural data scanning range

[0056] Specifically, in Step 121, the selected data set is divided into a training set, a validation set, and a test set in a ratio of 80%, 10%, and 10%.

[0057] Further, in Step 122, Stores the predicted spectra generated by the GLSAT forward prediction network.

[0058] In Step 151, considering that the reflection spectrum usually presents a Gaussian line shape and the spectral bandwidth of high reflectivity, the standard deviation is further defined , parameterize the Gaussian curve to construct a set of Gaussian-like reflectance spectrum data sets as Gaussian-type spectra; use represents the Gaussian spectrum obtained by GSSG processing.

[0059] Furthermore, for a single-peak spectrum, its Gaussian spectrum curve function is shown in formula (11), and for a double-peak spectrum, its Gaussian spectrum curve function is shown in formula (12): (11) (12) In the process of training the DNN inverse network in Step 152, in view of the key importance of the reflection peak in the design of the metasurface filter, a weight factor is introduced. W S , to enhance the impact of peak features on prediction results, so that the network can preferentially capture key spectral characteristics. W S It is given by the following formula (13): (13) It should be noted that for other application scenarios, such as broadband filters and transmission filters, the weighting factors can be adjusted or even the loss function can be changed to meet specific ideal spectral response requirements.

[0060] Furthermore, by minimizing the Euclidean distance between the predicted spectrum and the input Gaussian spectrum data, the weights and biases of each linear layer in the DNN are iteratively updated, while the network parameters of the pre-trained GLSAT forward prediction network remain unchanged. As the training proceeds, the network parameters in the DNN are gradually optimized until an optimized DNN reverse design network is obtained.

[0061] Preferably, the DNN is trained using a loss function defined by the following formula (14): , (14) in, Represents the number of training set samples of the DNN reverse design network, represents the resonant metasurface structure parameter value generated by the DNN inverse design network, = represents the resonant metasurface structure parameter value selected in Step 120. It can be seen that the loss function It is composed of two parts: structural parameter loss and spectral data loss, which are weighted by a weight factor to combine.

[0062] During the training and optimization of the DNN, the optimization performance index is defined by the following formula: , (15) The indicators The validity and reliability of the model are ensured by rigorously evaluating the accuracy of reproducing the ideal spectrum.

[0063] Finally, through the Step 160, the optimized DNN reverse design network can be used to generate the resonant metasurface structure parameters with high reflectivity in the target band on demand at one time.

[0064] The simulation of the second embodiment was carried out, and the Lumerical FDTD numerical simulation software was used to generate the original spectrum in the wavelength range of 700-1100 nm, which contained 300 equidistant points of wavelength. Finally, 2120 sets of sample data were obtained, and the samples included high reflectivity spectra with single-band reflection peaks or double-band reflection peaks. The algorithm effect verification of all subsequent networks was carried out based on this data set.

[0065] In order to verify the accuracy of the data-driven cascade inverse design network, the present invention uses 212 test data samples that have not been used in the data set. The difference of each spectral point is again measured by the absolute value error, and the average value is only 0.0063. The results of comparing the original spectrum obtained by simulation with the predicted spectrum are shown in Figure 2. Figure 8 The difference between the two is quantified by the different mean absolute errors (MAE) between the spectra and marked in Figure 8 Above each comparison graph in (a)-(h), the absolute value error at each spectral point is represented by the shaded area in the graph, further demonstrating the high accuracy of the GLSAT forward training network. Fig. 9 What is displayed is the distribution histogram of the absolute value error range of each spectral point, which once again illustrates that the cascaded inverse design network designed in the method described in the present invention can achieve the high-precision prediction requirements of the ideal spectral response, thereby fully proving the high performance of the method steps in realizing the resonant metasurface design based on spectral perception improvement.

[0066] In order to further illustrate the forward prediction effect of the GLSAT forward prediction network designed by the present invention, simulation and comparison are performed on multiple hypersurface design data sets in the prior art. The comparative data of the GLSAT forward prediction network and the prior art are shown in Table 2: Table 2. Experimental comparison data of GLSAT forward prediction network and existing technology

[0067] Among them: the data columns where the diamond reflector and silicon carbide reflector of the data set are the data of experiments conducted using the technical solution of the present invention, and the data columns corresponding to the ultrasonic absorber, four-resonator metasurface, nanosphere particles, silicon cylindrical metasurface and H-type silicon cylindrical metasurface of the data set are all the data of experiments conducted using the existing technology.

[0068] Table 2 shows the forward prediction results of the GLSAT forward prediction network in the method described in the second embodiment on different data sets provided by the present invention and the above-mentioned prior art. The dimensions of the resonant metasurface structural parameters corresponding to these data sets are at least 3 and at most 2002, and the number of samples is at least 2021 and at most 200000. In the improved resonant metasurface design method based on spectral perception provided by the present invention, the algorithm framework of the designed GLSAT forward prediction network can basically achieve high-precision prediction on these data sets, which shows that the GLSAT forward prediction network has a strong generalization ability.

[0069] It can be seen from the last row in the table that, except for the silicon cylindrical metasurface dataset in the prior art, which uses a downsampling process of the spectral dimension from 301 to 31, resulting in the loss of spectral detail information and cannot be compared, the GLSAT forward prediction network designed in the resonant metasurface design method provided by the present invention has achieved improved performance indicators on all other prior art datasets in Table 2. This shows that the GLSAT forward prediction network designed by the present invention has a higher ideal spectral prediction accuracy.

[0070] In Step 160 of the second embodiment, the resonant supersurface structure parameters of the optimized DNN reverse design network are obtained, and the corresponding predicted spectrum is obtained through the optimized GLSAT forward prediction network. , and the The Gaussian spectrum data generated by GSSG at the beginning The comparison results are as follows Fig.10 In addition, to further verify the effectiveness of the method described in the present invention, the resonant metasurface structural parameters predicted by the DNN inverse design network are used, and the corresponding simulation spectrum is calculated again through FDTD simulation. , and the original spectrum before GSSG processing The comparison results are as follows Fig.10 The results show that the DNN inverse design network designed in the method of the present invention can effectively predict the structural parameters of the resonant metasurface that meets the ideal spectral response requirements, and the predicted spectrum is highly consistent with the ideal spectral response requirements of the narrow-band filtering reflective resonant metasurface.

[0071] In a third embodiment, the present invention provides a resonant metasurface design device based on spectral perception improvement, wherein the device designs a resonant metasurface using the steps of the aforementioned method, and the device includes the following modules: The first module is used to design a GLSAT forward prediction network to achieve forward prediction from the resonant metasurface structure parameters to the ideal spectrum and obtain the predicted spectrum; the design of the GLSAT forward prediction network is based on the multi-head attention mechanism improved by spectral perception; the GLSAT forward prediction network includes an FCL module for dimensional expansion and a dual Transformer module for feature decomposition and extraction of high-dimensional spectra; The second module is used to select a data set of resonant metasurface structure parameters according to the predicted design requirements for achieving an ideal spectral response, train and optimize the GLSAT forward prediction network, and obtain a trained GLSAT forward prediction network; The third module is used to design a DNN reverse design network to achieve the design of resonant metasurface structure parameters that meet the ideal spectral response; the DNN reverse design network implements the design algorithm based on a deep neural network; The fourth module is used to cascade the DNN reverse design network with the trained GLSAT forward prediction network to construct a cascade reverse design network from the ideal spectrum to the resonant metasurface structure parameters; the cascade reverse design network includes a Gaussian line spectrum generator for preprocessing the ideal spectrum, a DNN reverse design network and a trained GLSAT forward prediction network; The fifth module is used to input the ideal spectrum as a training set into the cascade inverse design network, train and optimize the DNN inverse design network, until a cascade inverse design network that meets the design accuracy requirements is obtained; The sixth module is used to input a set of ideal spectra based on actual design requirements into the cascaded inverse design network that meets the design accuracy requirements, and use the DNN inverse design network in the cascaded inverse design network to generate resonant metasurface structural parameters that meet the ideal spectral response, thereby completing the design of the resonant metasurface.

[0072] Compared with the prior art, the present invention provides a resonant metasurface design method and device based on improved spectral perception, which combines global and local spectral perception capabilities to effectively solve the information loss problem in high-dimensional spectral prediction from low-dimensional structural data; reduce the complexity and computational cost of the model, and improve the accuracy and efficiency of spectral prediction and reverse design. Another technical improvement is that higher prediction accuracy can be achieved under the premise of a small number of data set samples.

[0073] The present invention also provides a computer device in one embodiment, including a memory and a processor, the memory stores a computer program, and the processor implements the steps of the improved resonant metasurface design method based on spectral perception provided in any of the above embodiments when executing the computer program. The computer device can be a server. The computer device includes a processor, a memory, a network interface and a database connected via a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store sample data. The network interface of the computer device is used to communicate with an external terminal through a network connection.

[0074] On the other hand, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the resonant metasurface design method based on spectral perception improvement provided in any of the above embodiments.

[0075] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0076] Matters not covered by the present invention are known technologies.

[0077] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0078] The above-described embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present application shall be subject to the attached claims.

[0079] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A resonant metasurface design method based on spectral perception improvement, characterized in that: The method comprises the following steps: Step 110: Design a GLSAT forward prediction network to achieve forward prediction from resonant metasurface structural parameters to ideal spectra to obtain predicted spectra; the design of the GLSAT forward prediction network is based on a multi-head attention mechanism improved by spectral perception; the GLSAT forward prediction network includes an FCL module for dimensional expansion and a dual Transformer module for feature decomposition and extraction of high-dimensional spectra; Step 120: According to the prediction design requirements for achieving an ideal spectral response, a data set of resonant metasurface structure parameters is selected, and the GLSAT forward prediction network is trained and optimized to obtain a trained GLSAT forward prediction network; Step 130: Design a DNN reverse design network to achieve the design of resonant metasurface structure parameters that meet the ideal spectral response; the DNN reverse design network implements the design algorithm based on a deep neural network; Step 140: Cascading the DNN reverse design network with the trained GLSAT forward prediction network to construct a cascade reverse design network from an ideal spectrum to resonant metasurface structural parameters; the cascade reverse design network includes a Gaussian line spectrum generator for preprocessing the ideal spectrum, a DNN reverse design network, and a trained GLSAT forward prediction network; Step 150: inputting the ideal spectrum as a training set into the cascade inverse design network, training and optimizing the DNN inverse design network, until a cascade inverse design network that meets the design accuracy requirements is obtained; Step 160: Input a set of ideal spectra based on actual design requirements into the cascaded inverse design network that meets the design accuracy requirements, and use the DNN inverse design network in the cascaded inverse design network to generate resonant metasurface structural parameters that meet the ideal spectral response, thereby completing the design of the resonant metasurface.

2. The resonant metasurface design method based on spectral perception improvement according to claim 1 is characterized in that: The FCL module includes multiple residual modules with a fully connected layer structure, which are used to achieve three-stage dimensional expansion. The fully connected layer structure includes multiple linear layers; the residual modules include at least three types: a first residual module, a second residual module and a third residual module; each linear layer of any residual module is followed by an activation function LeakyReLU, which is used to introduce nonlinear factors and alleviate gradient disappearance, and its negative slope is ; The dual Transformer module is used to extract global and local features of high-dimensional spectral data, and extract all features through a multi-head attention mechanism; the dual Transformer module includes a GPT module for global perception and an LPT module for local perception; the multi-head attention mechanism includes a cross-attention mechanism for the GPT module to extract global features to realize interactive multi-head attention analysis of spectral fragments, and a self-attention mechanism for the LPT module to extract local features to realize inline multi-head attention analysis of spectral fragments.

3. The resonant metasurface design method based on spectral perception improvement according to claim 2 is characterized in that: In Step 110, the GLSAT forward prediction network is used to achieve forward prediction from the resonant metasurface structure parameters to the ideal spectrum, and the process of obtaining the predicted spectrum includes: Step 111: Dimensionally expand the resonant metasurface structure parameters using the FCL module, including: The parameters of the dimensional resonant metasurface structure are expanded to The two first residual modules of represents the number of resonant metasurface structure parameters; the first residual module is expanded to The second residual module is further expanded to dimensional third residual module, where represents the spectral dimension; Step 112: preprocessing the output data of the FCL module before entering the dual Transformer module, including: using a layer of dimension The fully connected layer structure is used to standardize the data. is the standardized spectral dimension; the standardized data is processed through the projection matrix to obtain the query vector required by the multi-head attention mechanism , key vector Sum value vector ,in , is the total number of tokens in the spectrum, and token represents a spectrum fragment; Step 113: input the data preprocessed by Step 112 into the dual Transformer module for data processing, including: data processing by the GPT module, data processing by the LPT module, and smoothing the data by a one-dimensional convolution layer connecting the GPT module and the LPT module, and the LPT module and the data output layer respectively; Step 114: The data smoothed by the one-dimensional convolution layer connecting the LPT module and the data output layer is output through the output layer to obtain a predicted spectrum.

4. The resonant metasurface design method based on spectral perception improvement according to claim 3 is characterized in that: In Step 113, The data processing process of the GPT module includes: Divide the data into Tokens; Use cross attention mechanism to capture the attention of token interaction, attention score and attention results Calculated by the following formulas: in, Representation Matrix The transpose of Indicates the transpose of a matrix, and the Softmax function is used to convert a matrix data into a probability value score; By focusing all the heads Splicing together to obtain the multi-head attention analysis result based on the cross-attention mechanism is given by the following formula: Among them, the Contact function is a tensor concatenation function used to achieve multiple attention results splicing; It is The attention result of heads, the number of heads is , is the weight parameter matrix associated with all heads; The obtained multi-head attention analysis results , passed to a normalized feed-forward layer to implement multi-layer normalization processing of data; The data processing process of the LPT module includes: Use Contains The self-attention mechanism of each attention head will perform multi-head attention analysis on a single token to obtain an attention score , given by the following formula: in, ,all and The data are all from the same spectral fragment. The Sigmoid function is an S-shaped activation function, and its output value is between; use and The Hadamard product of the attention result is obtained , is given by: Among them, all , and The data are all derived from the same spectral fragment; Use the Contact function to collect the attention results of all heads Putting them together, we get the multi-head attention analysis result based on the self-attention mechanism, which is given by the following formula: in, Indicates the number of heads, It is The attention result of each head, is the weight parameter matrix associated with all heads; The obtained multi-head attention analysis results , passed to a normalized feed-forward layer to implement multi-layer normalization processing of data.

5. The resonant metasurface design method based on spectral perception improvement according to claim 4 is characterized in that: The normalized feedforward layer includes two normalization layers, a feedforward network, and an activation function LeakyReLU; the two normalization layers are the first normalization layer and the second normalization layer; the structure of the normalized feedforward layer from top to bottom is: the first normalization layer is used to ensure the stability of feature distribution and improve the convergence performance of the model; the feedforward network is used to perform nonlinear transformation and further extract features; The second normalization layer is used for stable gradient propagation; the activation function LeakyReLU is used to introduce nonlinear factors to prevent gradient disappearance; During the multi-layer normalization processing of the data, after the data is transmitted from the first normalization layer to the feedforward network, the input and output of the feedforward network are fused by a residual connection to ensure that the input information of the data is retained.

6. The resonant metasurface design method based on spectral perception improvement according to claim 5 is characterized in that: In step 120, according to the prediction design requirements for achieving an ideal spectral response, a data set of resonant metasurface structure parameters is selected to train and optimize the GLSAT forward prediction network, including: Step 121: Divide the data set into a training set, a validation set, and a test set in proportion; Step 122: The GLSAT forward prediction network is trained and optimized using the AdamW optimizer. The initial value of the learning rate in the AdamW optimizer is set to 0.0005. The learning rate is adjusted by cosine annealing scheduling and gradually decreases according to the cosine function. An L2 regularization function module is added to the AdamW optimizer to reduce overfitting. The loss function in the AdamW optimizer is The mean square error is used to measure the predicted spectrum. Compared with the original spectrum obtained by simulation The difference between the loss function Given by: ， in, Indicates the number of samples in the training set; Step 123: Through the indicators on the validation set , evaluate the optimization effect of the GLSAT forward prediction network; the index The mean absolute error is calculated to provide a robust evaluation result of network optimization performance. Given by: ， in, Indicates the number of samples in the validation set.

7. The method for designing a resonant metasurface based on improved spectral perception according to claim 6, characterized in that: In Step 130, the input layer of the DNN reverse design network contains spectral data, spectral points, the output layer generates a vector, storing the corresponding data of the resonant supersurface structure parameters; the deep neural network is composed of three consecutive fully connected hidden layers with a fully connected layer structure, and the three fully connected hidden layers respectively contain , and A neuron.

8. The method for designing a resonant metasurface based on improved spectral perception according to claim 7, characterized in that: In step 150, the ideal spectrum as a training set is input into the cascade inverse design network, and the DNN inverse design network is trained and optimized until a cascade inverse design network that meets the design accuracy requirements is obtained, including: Step 151: preprocessing the ideal spectrum using the Gaussian line spectrum generator to obtain a Gaussian spectrum that is closer to a physically achievable spectrum, including constructing a set of Gaussian-like spectrum data sets to obtain a Gaussian spectrum corresponding to the response of the ideal spectrum; The ideal spectrum is obtained by filtering the original spectrum data set through a predefined requirement threshold to generate an ideal spectrum containing only 0 or 1 binarization; the screening threshold is set to 0.9; Step 152: Input the Gaussian spectrum into the cascade architecture of the DNN reverse design network and the trained GLSAT forward prediction network, and train and optimize the DNN reverse design network, including: inputting the Gaussian spectrum into the DNN reverse design network, using the DNN reverse design network to design, generate and output resonant supersurface structure parameters; using the output resonant supersurface structure parameters as input data of the trained GLSAT forward prediction network to generate and output a predicted spectrum; comparing the input Gaussian spectrum with the output predicted spectrum, calculating the loss function, and optimizing the parameters of the deep neural network by minimizing the gradient descent of the loss function, thereby improving the accuracy of the DNN reverse design network in designing resonant supersurface structure parameters.

9. The method for designing a resonant metasurface based on improved spectral perception according to claim 8, characterized in that: The predicted design requirements of the ideal spectrum meet the requirements of the designed resonant metasurface on the wavelength and intensity of the reflection peak, including a narrow-band filtering reflection resonant metasurface specially designed for an alkali metal laser with a line width of 10 nm and a wavelength of 795 nm; The data set selected in Step 120 is the resonant supersurface structure parameters of the square lattice diamond structure silicon dioxide substrate, including: the diameter of the cylinder ,high The distance between adjacent cylinders , where the diameter Less than distance , the dimension of the resonant metasurface structure parameters is , set the training rounds ; In the Step 121, the data set is divided into a training set, a validation set, and a test set according to a ratio of 80%, 10%, and 10%; In Step 151, by defining the standard deviation , parameterize the Gaussian line-shaped reflectance spectrum curve and construct a set of Gaussian-like spectrum data sets as Gaussian spectrum; represents the Gaussian spectrum obtained by the Gaussian line shape processor; The process of training the DNN reverse design network in Step 152 includes: Iteratively updating the weights and biases of each fully connected hidden layer in the deep neural network by minimizing the Euclidean distance between the input Gaussian spectrum and the predicted spectrum generated by the trained GLSAT forward prediction network, while keeping the network parameters of the trained GLSAT forward prediction network unchanged, and gradually optimizing the network parameters in the deep neural network until an optimized DNN reverse design network is obtained; Introducing weight factors W S , to enhance the impact of peak features on prediction results, so that the network can preferentially capture key spectral characteristics. W S Given by: ； The DNN inverse design network is trained using a loss function defined by the following formula : ， in, Represents the number of training set samples of the DNN reverse design network, represents the resonant metasurface structure parameters generated by the DNN inverse design network, = represents the resonant metasurface structure parameter value selected in Step 120; The optimization performance index of the DNN reverse design network is defined by the following formula: ； In Step 160, the optimized DNN reverse design network is used to generate the resonant metasurface structure parameters with high reflectivity in the target band at one time according to the predicted design requirements of the ideal spectrum.

10. A resonant metasurface design device based on spectral perception improvement, characterized in that: The device designs a resonant metasurface using the steps of the method according to claim 1, and the device comprises the following modules: The first module is used to design a GLSAT forward prediction network to achieve forward prediction from the resonant metasurface structure parameters to the ideal spectrum and obtain the predicted spectrum; the design of the GLSAT forward prediction network is based on the multi-head attention mechanism improved by spectral perception; the GLSAT forward prediction network includes an FCL module for dimensional expansion and a dual Transformer module for feature decomposition and extraction of high-dimensional spectra; The second module is used to select a data set of resonant metasurface structure parameters according to the predicted design requirements for achieving an ideal spectral response, train and optimize the GLSAT forward prediction network, and obtain a trained GLSAT forward prediction network; The third module is used to design a DNN reverse design network to achieve the design of resonant metasurface structure parameters that meet the ideal spectral response; the DNN reverse design network implements the design algorithm based on a deep neural network; The fourth module is used to cascade the DNN reverse design network with the trained GLSAT forward prediction network to construct a cascade reverse design network from the ideal spectrum to the resonant metasurface structure parameters; the cascade reverse design network includes a Gaussian line spectrum generator for preprocessing the ideal spectrum, a DNN reverse design network and a trained GLSAT forward prediction network; The fifth module is used to input the ideal spectrum as a training set into the cascade inverse design network, train and optimize the DNN inverse design network, until a cascade inverse design network that meets the design accuracy requirements is obtained; The sixth module is used to input a set of ideal spectra based on actual design requirements into the cascaded inverse design network that meets the design accuracy requirements, and use the DNN inverse design network in the cascaded inverse design network to generate resonant metasurface structural parameters that meet the ideal spectral response, thereby completing the design of the resonant metasurface.

Images