Multifunctional metasurface optimization design method based on PF-LSTM-XGBoost

Abstract

The invention discloses a multifunctional metasurface optimization design method based on PF-LSTM-XGBoost, and belongs to the field of metamaterial and metasurface design, and the method comprises the steps: generating a metasurface unit structure and a data set of frequency response through CST-Python joint simulation; according to the data set, a PF-LSTM-XGBoost model is constructed and trained; and an RBMO optimization algorithm module is combined with a PF-LSTM-XGBoost model, and reverse design of a metasurface unit structure is carried out. In order to solve the design problem of various metasurfaces, the frequency response is predicted by adopting a mode of combining the PF-LSTM and the XGBoost model, so that the prediction precision of the frequency response is improved. Meanwhile, when the reverse design problem in the metasurface design is solved, a red-mouth blue-magpie optimization algorithm is introduced. And the adaptive capacity of the optimization process in various environments is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of metamaterial and metasurface design, and more particularly to a multifunctional metasurface optimization design method based on PF-LSTM-XGBoost. Background Art

[0002] As an artificial design means for regulating electromagnetic waves in terms of absorption, transmission, reflection, etc., metasurfaces have a wide range of applications in the fields of electronics, military, protection, and communication, and are particularly important in radar stealth, wireless communication, and signal shielding. With the rapid development of smart home, Internet of Things, and unmanned driving technologies, microwave transmitting and receiving devices have been widely integrated into daily life. At the same time, the complex electromagnetic environment has received increasing attention. How to effectively select and regulate electromagnetic signals in specific frequency bands has become a hot topic in the scientific research field.

[0003] Due to its compact size, metasurfaces can be more conveniently integrated into various systems, and have a high degree of design flexibility. The shape, size, arrangement, and material of the unit can be precisely adjusted to achieve precise regulation of electromagnetic waves. Through microstructural design, metasurfaces can achieve various effects such as reflection, refraction, diffraction, and polarization regulation, and can also achieve multifunctional design in different frequency bands. However, the relationship between the metasurface structure and electromagnetic response at the sub-wavelength scale is very complex and difficult to accurately describe by simple mathematical formulas. Relying solely on experience and electromagnetic knowledge, combined with numerical simulation and trial-and-error methods for optimization, designing a metasurface usually requires hundreds to thousands of simulation calculations, which is both time-consuming and requires multiple iterations. Facing the huge design space, traditional optimization methods often have difficulty in efficiently finding the global optimal solution.

[0004] With the progress of computer hardware, deep learning technology has been significantly improved and has played an important role in solving many complex scientific problems. In recent years, deep learning has gradually been applied to the field of metasurface design. It designs the metasurface microstructure in a data-driven manner, avoiding a large number of mathematical formula calculations and the cumbersome parameter scanning and trial-and-error processes, and significantly improving the design efficiency.

[0005] Therefore, how to provide a multifunctional metasurface optimization design method based on PF-LSTM-XGBoost is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a multifunctional metasurface optimization design method based on PF-LSTM-XGBoost to solve the technical problems existing in the above-mentioned prior art.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A multifunctional metasurface optimization design method based on PF-LSTM-XGBoost, including:

[0009] Using CST-Python co-simulation to generate a dataset of metasurface unit structures and frequency responses;

[0010] Construct and train a PF-LSTM-XGBoost model according to the dataset;

[0011] Combine the RBMO optimization algorithm module with the PF-LSTM-XGBoost model to perform reverse design of the metasurface unit structure.

[0012] Furthermore, using CST-Python co-simulation to generate a dataset of metasurface unit structures and frequency responses, including:

[0013] Based on the MDM method, design the metasurface array structure through 0 / 1 coding, and set different structure coding matrices and dielectric thicknesses;

[0014] Use CST-Python co-simulation to generate metasurface frequency response data;

[0015] Combine the metasurface unit structure pattern, frequency response, unit period and dielectric thickness to generate a dataset of metasurface unit structures and frequency responses.

[0016] Furthermore, it also includes: dividing the dataset into a training set and a test set.

[0017] Furthermore, the PF-LSTM-XGBoost model is composed of a parallel connection of a PF-LSTM model and an XGBoost model, where:

[0018] The PF-LSTM model consists of a PF layer, an LSTM layer, a fully connected layer, and a Dropout layer;

[0019] The XGBoost model combines multiple weak learners together to obtain a strong learner for final output prediction of the frequency response;

[0020] Perform weighted averaging on the prediction results and output the final predicted value of the frequency response.

[0021] Furthermore, the expression of the PF-LSTM model is:

[0022] X∈R M×L ;

[0023]

[0024] In the formula, X is the input sequence data, containing M samples, L is the length of each sample, denotes the Fourier transform, is the inverse Fourier transform, means that the planar shaping filter shares parameters among all channels, ⊙ L represents the element-wise product on the L-th dimension, that is, the element-wise multiplication operation.

[0025] Furthermore, the expression of the XGBoost model is:

[0026]

[0027] In the formula, the initial predicted value is the mean of the M sample labels, r i (1) is the difference between the true label of each sample and the current predicted value. The residual represents the error of the current prediction of the model, y i represents the true value of the i-th sample, is the predicted value of the previous iteration, η is the learning rate, f t (X) is the prediction result of the t-th tree, representing the output given by the tree in the current feature space, is the final predicted result.

[0028] Furthermore, the expression of the weighted average is:

[0029]

[0030]

[0031] P c = P l × α + P z × β;

[0032] In the formula, α is the weight of the prediction result of the PF-LSTM model, β is the weight of the prediction result of the XGBoost model, E is the error between the prediction result of the PF-LSTM model and the true value, Z is the error between the prediction result of the XGBoost model and the true value, P c is the predicted frequency response output by the PF-LSTM-XGBoost prediction network model, P l is the predicted frequency response output by the PF-LSTM model, P z is the predicted frequency response output by the XGBoost model.

[0033] Furthermore, the mean squared error MSE is used as the loss function of the PF-LSTM-XGBoost model, and the expression is:

[0034]

[0035] Among them, Tr is the true value.

[0036] Further, the RBMO optimization algorithm is combined with the PF-LSTM-XGBoost model for inverse design of the metasurface unit structure, including:

[0037] Input a desired frequency response target;

[0038] Enter the RBMO optimization algorithm module. After performing search, attack, and storage operations, generate preliminary metasurface data;

[0039] Use the PF-LSTM-XGBoost model to process the preliminarily generated metasurface data and predict the actual frequency response of the metasurface;

[0040] Judge whether the actual frequency response meets the designed frequency response. If the predicted actual frequency response does not meet the designed frequency response, return to the RBMO optimization algorithm module to perform search, attack, and storage operations again until a suitable metasurface structure is found;

[0041] If the predicted actual frequency response meets the requirements of the designed frequency response, output the finally determined encoding structure of the metasurface unit.

[0042] It can be seen from the above technical solutions that compared with the prior art, the present invention discloses a multifunctional metasurface optimization design method based on PF-LSTM-XGBoost, which solves the problems that the existing forward prediction model of metasurfaces is difficult to fit complex frequency responses and the traditional inverse design method is inefficient. By combining the PF-LSTM-XGBoost prediction model and the optimization algorithm, it effectively ensures the efficient training of the forward prediction network and the inverse design network, improves the accuracy and efficiency of forward frequency response prediction, and automatically selects the optimal design during optimization. By constructing a training dataset containing frequency response requirements and microstructural design parameters, the deep neural network can quickly predict and design the microstructural patterns and structural parameters of the metasurface without a large amount of simulation calculations. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to the provided drawings.

[0044] Figure 1 Schematic diagram of the encoded metasurface unit and the metasurface constituted thereby provided by the present invention;

[0045] Figure 2The flow chart for collecting the CST-Python co-simulation data set provided by the present invention;

[0046] Figure 3 The forward prediction flow chart of the frequency response of the PF-LSTM-XGBoost model provided by the present invention;

[0047] Figure 4 The schematic diagram of the PF-LSTM model structure provided by the present invention;

[0048] Figure 5 The flow chart of the prediction model fusion RBMO optimization algorithm provided by the present invention;

[0049] Figure 6 The comparison chart of the input and output results of the absorption rate and transmission coefficient provided by the present invention;

[0050] Figure 7 The schematic flow chart of the multi-functional metasurface optimization design method based on PF-LSTM-XGBoost provided by the present invention. Detailed implementation manners

[0051] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0052] See Figure 7 As shown, the embodiments of the present invention disclose a multi-functional metasurface optimization design method based on PF-LSTM-XGBoost, including:

[0053] Using CST-Python co-simulation to generate a data set of metasurface unit structures and frequency responses;

[0054] According to the data set, construct and train a PF-LSTM-XGBoost model;

[0055] Combine the RBMO optimization algorithm module with the PF-LSTM-XGBoost model to perform reverse design of the metasurface unit structure.

[0056] In a specific embodiment, the main content of the present invention is to realize the rapid prediction of the metasurface frequency response through deep learning methods and to perform rapid design of the metasurface by establishing a suitable deep neural network, and to solve some problems existing in the current deep learning metasurface design field.

[0057] Specifically, the present invention first designs the metasurface array structure through 0 / 1 coding based on MDM (Metasurface Design Method), sets different structural coding matrices and medium thicknesses, and uses CST-Python co-simulation to generate metasurface frequency response data. Combine information such as the metasurface unit structure pattern, frequency response, unit period, and medium thickness to construct a diverse and rich unit structure dataset. The dataset contains the 0 / 1 coding matrix corresponding to different metasurface unit structure patterns and their corresponding frequency responses, which are divided into a training set and a test set to establish a foundation for the effective training of subsequent deep learning models. Then, by constructing a prediction model, deeply explore the relationship between the matrix data corresponding to the metasurface unit structure and the frequency response, especially providing more information in terms of the integrity and local dependence of the unit structure, enabling the model to more accurately fit the frequency response of the metasurface and greatly improving the prediction accuracy. Finally, use the method of combining an optimization algorithm with the prediction model for the inverse design of the metasurface. The model can select the most suitable metasurface unit structure pattern so that it can work under the specified target frequency response, and complete various artificial manipulation functions of electromagnetic waves such as absorption and transmission, thereby realizing efficient and automated inverse design of the metasurface. To implement the above method, the solution includes the following three parts:

[0058] Part 1: Use CST-Python co-simulation to generate metasurface unit structure and frequency response datasets.

[0059] First, use a 0 / 1 coding matrix to mathematically represent the metasurface unit structure. Randomly generate 0 / 1 matrix codes in Python, and start calling the underlying module of the electromagnetic simulation software CST in Python through the interface function to set global parameters, material parameters, and the metasurface unit structure converted from the code. Then, set the periodic structure boundary conditions and the solver to obtain the frequency response parameters of the metasurface composed of the periodic arrangement of the unit structure through the simulation of a single unit. Select the corresponding parameter results and save them to a folder according to the required metasurface function (transmission or absorption). The entire automated modeling process continues n times until the set upper limit of the number of times is reached. Figure 1 Schematic diagram of the working of the double-layer coded metasurface unit and the metasurface composed of it. The dataset collection flow chart is as Figure 2 shown.

[0060] Part 2: Construct a forward prediction model for the metasurface unit frequency response based on PF-LSTM-XGBoost.

[0061] Specifically, the PF-LSTM-XGBoost prediction model is composed of a PF-LSTM model and an XGBoost model in parallel. The prediction results of the two models are weighted and averaged, and the model with a smaller variance will be assigned a higher weight. The structure of the PF-LSTM-XGBoost prediction model is shown in Figure 3 as follows.

[0062] Specifically, step one: construct the PF-LSTM prediction model. For the prediction problem in metasurface design, an improved long short-term memory model (LSTM) is proposed, which integrates instance normalization and a frequency filtering block (PaiFilter, PF). This model can effectively solve the non-stationarity in the data and complex input dependencies, thereby improving the prediction accuracy and the generalization ability of the model. The PF-LSTM model consists of a PF layer, an LSTM layer, a fully connected layer, and a Dropout layer. The PF layer can approximately replace linear mapping and attention mapping and has excellent capabilities in dealing with high-frequency noise. The LSTM layer receives the input sequence and generates hidden states after processing. The fully connected layer maps the output of the LSTM to the target output space. The Dropout layer is used to prevent the model from overfitting. By randomly discarding some neurons during the training process, the generalization ability of the model is enhanced. The structure of the PF-LSTM model is shown in Figure 4 as follows.

[0063] where m is the number of input features and n is the number of predicted frequency responses. PF is a prediction module based on the Fourier transform, mainly used for data feature extraction, convolution, and prediction. In the PF layer, the frequency filter is designed and applied through the Fourier transform. The core idea is to enhance the model's learning ability of data features by filtering the input signal in the frequency domain. This method instantiates the frequency filter by randomly initializing learnable parameters and then applies these filters to the input data to achieve the enhancement of frequency domain features. The parameter sharing strategy between different channels is different. By combining this design, the filters of different channels can learn independently. The related formula is as follows:

[0064] X ∈ R M×L (1)

[0065]

[0066] X is the input sequence data, containing M samples, and the length of each sample is L. represents the Fourier transform. is the inverse Fourier transform. is the planar shaping filter that shares parameters among all channels. ⊙ L represents the element-wise product on the L-th dimension, that is, the element-wise multiplication operation.

[0067] Step 2: Construct an Extreme Gradient Boosting (XGBoost) prediction model. The XGBoost prediction model combines multiple weak learners to build a strong learner for prediction and finally outputs the predicted frequency response. XGBoost uses the gradient boosting tree algorithm, which gradually optimizes the prediction results through a series of regression trees. The initial predicted value is usually the mean of all sample labels. In the first iteration, the residual of each sample (i.e., the difference between the true value and the current predicted value) is calculated, and then the first regression tree is trained to fit these residuals. Each tree attempts to improve the model's prediction by fitting the residuals. The above process continues until the maximum number of trees is reached, and the prediction is continuously improved by calculating the residuals and training new trees in each round. After all iterations are completed, the final predicted value of the model is the weighted sum of the predicted values of all trees, and the related formula is as follows:

[0068]

[0069] where the initial predicted value is the mean of the labels of M samples. r i (1) is the difference between the true label of each sample and the current predicted value, and the residual represents the error of the current prediction of the model. y i represents the true value of the i-th sample, is the predicted value of the previous iteration. η is the learning rate, which controls the magnitude of the update in each round. f t (X) is the prediction result of the t-th tree, indicating the output given by the tree in the current feature space. The final prediction result is the weighted sum of the prediction results of all trees.

[0070] Step 3: Perform weighted averaging on the prediction results to output a more accurate predicted frequency response value. The prediction results of these two models are weighted. The rule for calculating the weights is that the model with a smaller variance will be given a higher weight. The formula for calculating the weighting coefficient is as follows:

[0071]

[0072] α is the weight of the prediction result of the PF-LSTM model, β is the weight of the prediction result of the XGBoost model, E is the error between the prediction result of the PF-LSTM model and the true value, and Z is the error between the prediction result of the XGBoost model and the true value. The formula for the combined prediction result is as follows:

[0073] P c = P l ×α + P z ×β (11)

[0074] P c is the predicted frequency response output by the PF-LSTM-XGBoost prediction network model, P l is the predicted frequency response output by the PF-LSTM model, P z is the predicted frequency response output by the XGBoost model. The loss function is the mean squared error MSE, and its formula is as follows:

[0075]

[0076] where Tr is the true value.

[0077] Part Three: Use the RBMO optimization algorithm for inverse design of the metasurface unit structure.

[0078] The metasurface design task usually involves complex interactions between multiple parameters. The Red-billed Blue Magpie Optimizer (RBMO) can play a significant advantage in such tasks. The RBMO algorithm has three stages, namely search, attack, and storage. By mimicking the cooperation of biological groups and individual behaviors, RBMO has a powerful global search ability, which is crucial for avoiding being trapped in local optimal solutions. The metasurface design faces complex non-linear relationships between multiple parameters, and there are often multiple local optimal solutions in the solution space, which makes the algorithm easily fall into a sub-optimal solution. Through the cooperation of multiple search individuals, RBMO can widely explore the solution space, avoid local optimal solutions, and thus improve the global search effect. The following formula is used for iteration during exploration.

[0079]

[0080] where t represents the current iteration number, Xi(t + 1) represents the position of the i-th new search agent, p represents the number of red-billed blue magpies in 2 to 5 small groups randomly selected from all search individuals, Xm represents the randomly selected m-th individual, Xi represents the i-th individual, and Xrs(t) represents the search agent randomly selected in the current iteration.

[0081] In the attack stage, RBMO can flexibly adjust the search direction under different search stages and strategies. By adjusting the search strategy, the algorithm can effectively optimize the metasurface structure parameters and help designers achieve better design goals. The algorithm individuals adopt various optimization strategies to accurately capture potential high-quality solutions in the solution space. For smaller sub-problems or solution space regions, the individuals will quickly adjust the search direction and accurately locate the optimal solution. These fast and flexible strategies ensure that the algorithm can efficiently converge to potential high-quality solutions. The mathematical model of this process can be described by the following formula.

[0082]

[0083] where X f (t) represents the position of the food, and CF represents a varying adjustment factor used to control the step size of individual updates. It adjusts the amplitude of individual position updates, thereby affecting the exploration and exploitation capabilities of the search process. Rand represents a random number used to generate a standard normal distribution, t represents the current iteration number, and T represents the total number of iterations.

[0084] The stability and robustness of the algorithm are crucial during the optimization process. The storage part continuously records and saves high-quality solutions, enabling the algorithm to recover to a better solution state when encountering difficulties or fluctuations, enhancing the algorithm's adaptability and robustness in various environments. This is particularly useful for handling various challenges that may arise during the metasurface design iteration. Its mathematical model is shown as follows.

[0085]

[0086] and represent the fitness values of the i-th red-billed blue magpie before and after position update, respectively. Finally, the overall structure of the prediction model integrated with the RBMO optimization algorithm is as Figure 5 shown.

[0087] Among them, the prediction model is trained as a mapper from the metasurface structure to the frequency response. For the input encoded structure, the network can easily output the corresponding frequency response. The trained network is used to replace the calculation process of the electromagnetic simulation software and is used to calculate the fitness value of individuals in the RBMO optimization algorithm, accelerating the iterative process of the algorithm. By reverse-designing the metasurface and comparing it with the results of electromagnetic simulation, the comparison results show the effectiveness and accuracy of the proposed method. Compared with traditional methods, the proposed method has the characteristics of fast design speed and high design efficiency.

[0088] The metasurface structure designed by the optimization model is input into the electromagnetic simulation software CST and compared with the results of electromagnetic simulation. The comparison results show the effectiveness and accuracy of the proposed method. Compared with traditional methods, the proposed method has the characteristics of fast design speed and high design accuracy. As Figure 6 shown, by inputting the corresponding parameter frequency response according to the requirements of different functions, the method of the proposed PF-LSTM-XGBoost prediction model combined with the RBMO optimization algorithm can well complete the on-demand design of the metasurface structure and achieve the preset electromagnetic wave manipulation function.

[0089] This solution aims to solve the design problems of multiple metasurfaces. It uses the combination of PF-LSTM and XGBoost models to predict the frequency response, integrating the prediction results of the two models to increase the generalization ability, thereby improving the prediction accuracy of the frequency response.

[0090] When solving the inverse design problem in metasurface design, this solution introduces the red-billed blue magpie optimization algorithm. This algorithm simulates the cooperation of biological groups and individual behaviors, has strong global search ability, can avoid falling into local optimal solutions, helps designers better optimize the structural parameters of metasurfaces, and ensures the adaptability of the optimization process in various environments.

[0091] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other.

[0092] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multifunctional supersurface optimization design method based on PF-LSTM-XGBoost, characterized in that: include: Generate a dataset of metasurface unit structures and frequency responses using CST-Python co-simulation; Based on the dataset, build and train the PF-LSTM-XGBoost model; The RBMO optimization algorithm module is combined with the PF-LSTM-XGBoost model to perform reverse design of the hypersurface unit structure.

2. A multifunctional supersurface optimization design method based on PF-LSTM-XGBoost according to claim 1, characterized in that: CST-Python co-simulation was used to generate a dataset of metasurface unit cell structures and frequency responses, including: Based on the MDM method, the metasurface array structure is designed through 0 / 1 coding, and different structural coding matrices and dielectric thicknesses are set; Generate metasurface frequency response data using CST-Python co-simulation; The metasurface unit structure pattern, frequency response, unit period, and dielectric thickness are combined to generate a dataset of metasurface unit structure and frequency response.

3. A multifunctional supersurface optimization design method based on PF-LSTM-XGBoost according to claim 2, characterized in that: Also includes: The dataset is divided into a training set and a test set.

4. The multifunctional supersurface optimization design method based on PF-LSTM-XGBoost according to claim 1, characterized in that: The PF-LSTM-XGBoost model is composed of a PF-LSTM model and an XGBoost model in parallel, wherein: The PF-LSTM model consists of a PF layer, an LSTM layer, a fully connected layer, and a Dropout layer; The XGBoost model combines multiple weak learners together to obtain a strong learner to finally output the predicted frequency response; The prediction results are weighted averaged and the final frequency response prediction value is output.

5. The multifunctional supersurface optimization design method based on PF-LSTM-XGBoost according to claim 4 is characterized in that: The expression of the PF-LSTM model is: X∈R M×L ; In the formula, X is the input sequence data, which contains M samples, and L is the length of each sample. represents the Fourier transform, is the inverse Fourier transform, is a planar shaping filter that shares parameters among all channels, L Represents the element-wise product in the Lth dimension, that is, an element-by-element multiplication operation.

6. A multifunctional supersurface optimization design method based on PF-LSTM-XGBoost according to claim 5, characterized in that: The expression of the XGBoost model is: In the formula, the initial prediction value is the mean of the M sample labels, r i (1) is the difference between the true label of each sample and the current predicted value. The residual represents the error of the current prediction of the model. i represents the true value of the i-th sample, is the predicted value of the previous iteration, η is the learning rate, and f t (X) is the prediction result of the tth tree, which represents the output given by the tree in the current feature space. is the final prediction result.

7. The multifunctional supersurface optimization design method based on PF-LSTM-XGBoost according to claim 6, characterized in that: The expression of the weighted average is: P c =P l ×α+P z ×β; In the formula, α is the weight of the prediction result of the PF-LSTM model, β is the weight of the prediction result of the XGBoost model, E is the error between the prediction result of the PF-LSTM model and the true value, Z is the error between the prediction result of the XGBoost model and the true value, and P c is the predicted frequency response output by the PF-LSTM-XGBoost prediction network model, P l is the predicted frequency response output by the PF-LSTM model, P z is the predicted frequency response output by the XGBoost model.

8. The multifunctional supersurface optimization design method based on PF-LSTM-XGBoost according to claim 7, characterized in that: The mean square error MSE is used as the loss function of the PF-LSTM-XGBoost model, and the expression is: Among them, Tr is the true value.

9. The multifunctional supersurface optimization design method based on PF-LSTM-XGBoost according to claim 1, characterized in that: The RBMO optimization algorithm is combined with the PF-LSTM-XGBoost model to perform reverse design of the hypersurface unit structure, including: Enter a desired frequency response target; Enter the RBMO optimization algorithm module, perform search, attack, and storage operations, and generate preliminary hypersurface data; The PF-LSTM-XGBoost model is used to process the initially generated metasurface data to predict the actual frequency response of the metasurface; Determine whether the actual frequency response meets the designed frequency response. If the predicted actual frequency response does not meet the designed frequency response, return to the RBMO optimization algorithm module and perform search, attack and storage operations again until a suitable metasurface structure is found. If the predicted actual frequency response meets the design frequency response requirements, the final metasurface unit coding structure is output.

Images