Forward Design Method and System for Wave Absorbing Metamaterials
By combining backpropagation neural networks and improved genetic algorithms, the design of microwave absorbing metamaterials was optimized, solving the problems of high computational cost and long cycle in existing technologies, and realizing fast and accurate broadband microwave absorber design.
Patent Information
- Application Number
- CN202510625776.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-15
- Publication Date
- 2026-11-17
AI Technical Summary
In the current technology, the design of efficient and broadband absorbers relies on human experience and manual design, which is computationally expensive and time-consuming, making it difficult to find a satisfactory design solution quickly.
A method combining backpropagation neural network and improved genetic algorithm was adopted to obtain high-quality samples through uniform sampling, random sampling and adding geometric space constraints, train the neural network model, and optimize the design of microwave absorbing metamaterial.
It significantly reduces computational costs and time, and can quickly and accurately find the optimal design solution in a multivariable parameter space, thereby improving design efficiency and accuracy.
Smart Images

Figure CN123120122A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of structural optimization design of microwave absorbing materials, and specifically to a forward design method and system for microwave absorbing metamaterials. Background Technology
[0002] With the rapid development of modern electronic information, wireless communication, and radar systems, the electromagnetic environment has become increasingly complex. Electromagnetic radiation not only has a serious impact on people's physical and mental health, but also adversely affects the operation of precision instruments and the improvement of equipment performance. Therefore, the protection against electromagnetic waves has become an urgent problem to be solved.
[0003] In existing technologies, the design of efficient, broadband absorbers typically relies on human experience and manual design. These methods require designers to possess profound professional knowledge, which severely limits the exploration of a vast design space, and often fails to find a satisfactory design solution after numerous attempts. While technological advancements have led to the development of theoretical methods, numerical simulations, reverse engineering, deep neural networks, and other forward and reverse design methods, these methods still require frequent calculations of the structure's absorption performance using the finite element method or other numerical simulation techniques. This results in excessively long calculation cycles, high computational costs, and significant application difficulties. Summary of the Invention
[0004] To address the above technical problems, this invention proposes a forward design method and system for absorbing metamaterials to improve their wave absorption performance.
[0005] The first aspect of this invention provides a forward design method for microwave absorbing metamaterials, comprising: Initial samples are obtained by uniformly sampling within the geometric space of the sample. The initial samples are then simulated at a preset frequency using preset simulation software to obtain the corresponding initial sample absorption performance parameters. These initial sample absorption performance parameters are then filled into the total dataset to train a preset neural network model. High-error samples are randomly sampled within the geometric space of the sample. The high-error samples are then simulated at a preset frequency using preset simulation software to obtain the high-error sample absorption performance parameters. Increase the constraints of the geometric space of the sample to obtain high-quality samples. Simulate the high-quality samples at a preset frequency using preset simulation software to obtain the high-quality sample absorption performance parameters of the corresponding high-quality samples. The absorption performance parameters of the high-error samples and the absorption performance parameters of the high-quality samples are respectively filled into the total dataset to gradually adjust the accuracy of the neural network model.
[0006] As an optional implementation, the preset neural network model is a backpropagation neural network model.
[0007] As an optional implementation, the step of uniformly sampling within the geometric space of the sample to obtain an initial sample, simulating the initial sample at a preset frequency using preset simulation software to obtain the corresponding initial sample absorption performance parameters, and filling the initial sample absorption performance parameters into the total dataset to train a preset neural network model includes: uniformly sampling within the geometric space of the sample to generate an initial sample using the Latin hypercube method. The corresponding absorption performance parameters are obtained by simulating at a preset frequency using electromagnetic field simulation software. The initial sample absorption performance parameters are filled into the total dataset to train the backpropagation neural network model.
[0008] As an optional implementation, the step of randomly sampling within the geometric space of the sample to obtain high-error samples includes: The individual x with the largest error compared to the simulation value is selected from the sample, and the k nearest neighbor samples x' of individual x are selected. nn ; Through x new =x + δ × (xx) nn Generate new samples x within a preset range of the individual region where the prediction error of the preset neural network model is the largest. new , as a high-error sample.
[0009] As an optional implementation, increasing the geometric space constraints of the samples to obtain high-quality samples includes: Using a genetic algorithm, the original population and corresponding individuals are configured. The binary string of each individual is divided into several parts according to the number of parameters, and each part represents a sample absorption performance parameter. Neighbor sampling was performed on the original population to fill in auxiliary populations; The elite selection strategy is used to select samples in the auxiliary population with the same number as the original population. Set the crossover probability and randomly generate the random probability; The crossover probability and random probability are compared, and parents are selectively obtained from the original population and auxiliary population to generate offspring populations, which are then used as high-quality samples.
[0010] As an optional implementation, the step of comparing crossover probabilities and random probabilities, selectively obtaining parent generations from the original population and auxiliary populations to generate offspring populations, and using the offspring populations as high-quality samples, includes: If the crossover probability is greater than the random probability, two individuals are randomly selected from the original population as parents. The two points intersect and fill the offspring population; If the crossover probability is not greater than the random probability, select one individual from the original population and one individual from the auxiliary population as the parent generation; The two points intersect and fill the offspring population.
[0011] As an optional implementation, the forward design method for absorbing metamaterials further includes obtaining and determining the number of random individuals in the offspring population based on the standard deviation factor of the population.
[0012] As an optional implementation, the forward design method for the absorbing metamaterial further includes using a penalty function to construct an evaluation function to obtain the lowest reflection loss within a specific frequency range.
[0013] As an optional implementation, the penalty function penalizes solutions that violate constraints by adding an extra term, wherein the penalty function is: The function f is an approximation of the reflectivity RL obtained by predicting through a pre-defined neural network model, where d, p, a, w, and r are the combined parameters of the absorbing material, and fq i Indicates frequency range, dB i θ is the reflectivity threshold at the i-th frequency, and θ is the penalty coefficient.
[0014] A second aspect of the present invention provides a forward design system for microwave absorbing metamaterials, comprising: A high-error sample acquisition unit is used at least to uniformly sample and acquire an initial sample in the geometric space of the sample, and to simulate the initial sample at a preset frequency using preset simulation software to obtain the corresponding initial sample absorption performance parameters. The high error sample acquisition unit is used to randomly sample and acquire high error samples in the geometric space of the sample, and to simulate the high error samples at a preset frequency using preset simulation software to obtain the high error sample absorption performance parameters of the corresponding high error samples. A high-quality sample acquisition unit is used to at least increase the geometric space constraints of the sample, acquire high-quality samples, and simulate the high-quality samples at a preset frequency using preset simulation software to obtain the high-quality sample absorption performance parameters of the corresponding high-quality samples. The model training and tuning unit is used at least to fill the initial sample absorption performance parameters into the total dataset to train a preset neural network model; the high-error sample absorption performance parameters and the high-quality sample absorption performance parameters are respectively filled into the total dataset to gradually adjust the accuracy of the trained neural network model. Attached Figure Description
[0015] Figure 1 This is a flowchart illustrating a forward design method for microwave absorbing metamaterials in one embodiment of the present invention.
[0016] Figure 2 This is a schematic diagram illustrating the operational principle of a forward design method for absorbing metamaterials according to an embodiment of the present invention.
[0017] Figure 3 This is a diagram illustrating the decimal to binary encoding of a microwave absorbing metamaterial according to an embodiment of the present invention.
[0018] Figure 4 This is a diagram showing the structural parameters of a BP neural network model in one embodiment of the present invention.
[0019] Figure 5 This is a graph showing the variation of reflectivity RL being less than -8GHz in the 2-18GHz range in one embodiment of the present invention.
[0020] Figure 6 This is a graph showing the variation of reflectivity RL being less than -10GHz in the 2-18GHz range in one embodiment of the present invention.
[0021] Figure 7 This is a distribution diagram of DE and Our-GA solutions for reflectivity RL less than -8GHz in the 2-18GHz range in one embodiment of the present invention.
[0022] Figure 8 This is a distribution diagram of DE and Our-GA solutions for reflectivity RL less than -8GHz in the 2-18GHz range in one embodiment of the present invention.
[0023] Figure 9 This is an average convergence curve of 100 iterations for each algorithm in the 2-18GHz range with a reflectivity RL less than -8GHz, as shown in one embodiment of the present invention.
[0024] Figure 10 This is an average convergence curve of 100 iterations for each algorithm in the 2-18GHz range with a reflectivity RL less than -8GHz, as shown in one embodiment of the present invention.
[0025] Figure 11 This is a block diagram of a forward design method for microwave absorbing metamaterials according to an embodiment of the present invention. Detailed Implementation
[0026] Genetic algorithms, heuristic algorithms proposed by Holland in 1975, are typically used to find the global maximum of a multidimensional objective function in a given parameter space. The application of artificial neural networks in the design of electromagnetic metamaterials overcomes the problem of requiring large amounts of computation in traditional methods. Furthermore, genetic algorithms (GA) can perform global searches, enabling them to explore a vast design parameter space. The combination of these two methods provides a novel approach to optimizing metamaterial structures, efficiently exploring the metamaterial design space and discovering structural parameters with optimal performance. This demonstrates high efficiency and accuracy in performance prediction, highlighting its enormous potential in the design and optimization of electromagnetic metamaterials.
[0027] For complex metamaterial structure design, researchers repeatedly call existing models to obtain the absorption performance of different designed structures. During the continuous design and adjustment of structural parameters, a large number of fitness assessments are required. Taking the iterative GA method as an example, assuming a population of 1000 individuals and a maximum generation count of 100, the entire process still requires 10⁵ simulations of absorption performance. Each simulation of a single sample takes approximately 3 minutes, and a complete experiment can take several months. To reduce the time cost of fitness assessment simulations, this paper uses a backpropagation neural network (BPNN or BP neural network) to evaluate individual fitness.
[0028] The present invention will be further illustrated by the following embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the present invention.
[0029] like Figure 1 and Figure 2 As shown, the first aspect of the present invention provides a forward design method for microwave absorbing metamaterials, comprising the following steps.
[0030] Step S100: Initial samples are uniformly sampled within the geometric space of the sample. The initial samples are simulated at a preset frequency using preset simulation software to obtain the corresponding initial sample absorption performance parameters. The initial sample absorption performance parameters are then filled into the total dataset to train a preset neural network model. Specifically, the preset neural network model is a backpropagation neural network model.
[0031] Step S200: Randomly sample high-error samples within the geometric space of the sample, and simulate the high-error samples at a preset frequency using preset simulation software to obtain the high-error sample absorption performance parameters corresponding to the high-error samples.
[0032] Step S300: Increase the geometric space constraints of the sample to obtain a high-quality sample. Simulate the high-quality sample at a preset frequency using preset simulation software to obtain the high-quality sample absorption performance parameters of the corresponding high-quality sample.
[0033] Step S400: Fill the total dataset with the absorption performance parameters of the high-error samples and the absorption performance parameters of the high-quality samples respectively, so as to gradually adjust the accuracy of the neural network model.
[0034] Specifically, in this invention, the dataset is also called the total dataset, denoted by M; the initial sample is also called the initial sample point, denoted by M0; the high error sample is also called the high error sample point, denoted by M1; and the high quality sample is also called the high quality sample point, denoted by M2. The acquisition of the dataset M is divided into three parts, including the initial sample M0, the high error sample M1, and the high quality sample M2.
[0035] like Figure 2 As shown, the present invention is generally divided into two parts: a data-driven generation module and an improved genetic algorithm module.
[0036] In one embodiment of the present invention, in step S100, the process of uniformly sampling within the geometric space of the sample to obtain an initial sample, simulating the initial sample at a preset frequency using preset simulation software to obtain corresponding initial sample absorption performance parameters, and filling the initial sample absorption performance parameters into the total dataset to train a preset neural network model includes: Step S110: Generate initial sample M0 by uniformly sampling within the geometric space of the sample using the Latin hypercube method; Step S120: Simulate the electromagnetic field within a preset frequency range using electromagnetic field simulation software to obtain the corresponding absorption performance parameters. For example, the preset frequency range listed in this invention is 2-18GHz. Step S130: Fill the initial sample absorption performance parameters into the total dataset M to train the backpropagation neural network (BP neural network) model.
[0037] In one embodiment of the present invention, in step S200, the step of randomly sampling within the geometric space of the sample to obtain high-error samples includes: Step S210: Select the individual x with the largest error from the simulation value in the sample M, and select the k nearest neighbor samples x of individual x. nn ; Step S220: via x new =x + δ × (xx) nn Generate a new sample x within a preset range δ of the individual region where the prediction error of the preset neural network model is the largest. new , as a high-error sample.
[0038] Specifically, in step S200, random sampling with replacement is performed within the geometric space of each variable. Duplicate sample points are removed, and these randomly sampled samples are simulated. A trained BPNN is then used to predict the absorption performance, with the reflection loss RL representing the absorption performance of the absorbing material at the target frequency. The mean square error (MSE) is used as the loss function. The individual x with the largest error from the simulation value is selected from the sample points, and the k nearest neighbor samples x' of individual x are chosen. nn Using x new =x + δ × (xx) nn Generate new samples x near the individual region with the largest BP prediction error. new M1 is a high-error sample.
[0039] In one embodiment of the present invention, in step S300, increasing the geometric space constraints of the samples to obtain high-quality samples includes: Using a genetic algorithm, the original population and corresponding individuals are configured. The binary string of each individual is divided into several parts according to the number of parameters, and each part represents a sample absorption performance parameter. Neighbor sampling was performed on the original population to fill in auxiliary populations; The elite selection strategy is used to select samples in the auxiliary population with the same number as the original population. Set the crossover probability and randomly generate the random probability; The crossover probability and random probability are compared, and parents are selectively obtained from the original population and auxiliary population to generate offspring populations, which are then used as high-quality samples.
[0040] In one embodiment of the present invention, the step of comparing the crossover probability and the random probability, selectively obtaining parent generations from the original population and the auxiliary population to generate offspring populations, and using the offspring populations as high-quality samples, includes: If the crossover probability is greater than the random probability, two individuals are randomly selected from the original population as parents. The two points intersect and fill the offspring population; If the crossover probability is not greater than the random probability, select one individual from the original population and one individual from the auxiliary population as the parent generation; The two points intersect and fill the offspring population.
[0041] For the electromagnetic performance optimization design of absorbing materials, especially multilayer materials, it is necessary to find a design scheme that meets the requirements over a wide frequency range. Traditional genetic algorithms have a simple structure, but they suffer from premature convergence and susceptibility to local optima. To address these shortcomings, this paper proposes an improved genetic algorithm. By constructing globally optimal individuals, guiding the population with auxiliary methods, and introducing new individuals, the diversity of the population is enhanced, enabling it to find the optimal design scheme under different design requirements.
[0042] Specifically, the genetic algorithm used in this invention is an improved genetic algorithm (Our-GA). By adding constraints to the geometric space, a high-quality sample M2 with wide bandwidth and high absorption rate is searched. The M1 and M2 pairs are simulated and filled into the dataset M, and the accuracy of the model is gradually adjusted. The specific overall process is as shown in Algorithm 1.
[0043] Specifically, this invention employs a binary encoding method, setting the encoding length according to the value range of different parameters, and concatenating them sequentially into a string. The population size is initialized to S, and S individuals are randomly generated. The binary string of each individual is divided according to the number of parameters, with each part representing a variable parameter of the absorber. The decimal to binary conversion process is as follows... Figure 3 As shown, each row of the binary array represents a set of parameter combinations for the absorber.
[0044] To increase population diversity and avoid getting trapped in local optima, this invention introduces auxiliary individuals, differing from previous methods in its selection and crossover operations. A two-way crossover approach is adopted, meaning that during the crossover process, an auxiliary population is constructed to guide individual populations iterating in directions with potential.
[0045] The pseudocode for the crossover process is shown in Algorithm 2: In the original population, firstly, neighbor sampling is performed on the individuals in the population, and then an elite selection strategy is used to select S individuals as auxiliary individuals. When performing the crossover operation, the crossover probability p is set. c Randomly generate p random and with crossover probability p c The two methods of crossover are compared to generate the offspring population. This crossover has two significant advantages: it utilizes the optimal individual information predicted by the BP network, accelerating the initial iteration speed, and it avoids the population getting trapped in local optima.
[0046] Algorithm 2 is specifically represented as follows:
[0047] Furthermore, although this invention predicts the reflection loss of the absorber (absorbing material) through a BP network, there is still a certain error in reality, and the obtained optimal solution may still be a local optimum. In order to increase the diversity of solutions, break the limitation of local optima, and obtain more design schemes, this invention also considers the strategy of introducing new individuals into the population. This helps to break the possible genetic rigidity. However, an excessive number of new individuals may dilute the beneficial genetic information in the population, thereby slowing down the convergence speed and search efficiency of the algorithm.
[0048] To balance this contradiction, in one embodiment of the present invention, a method is proposed to dynamically adjust the number of individuals introduced by obtaining and determining the number of random individuals in the offspring population based on the standard deviation factor of the population.
[0049] This invention provides a standard deviation factor as a statistical measure of the dispersion of genotype distribution, reflecting the level of gene diversity in the population. Let the population of generation t consist of N individuals, each containing L types of genes, i.e.:
[0050] in, This represents the l-th gene of the i-th individual in the t-th generation population. This represents the average number of individuals in the t-th generation population. The l-th gene represents the average individual in the t-th generation population: Therefore, the standard deviation of the population in generation t is: in: It is an L-dimensional row vector, which can be transformed into a one-dimensional variable: To represent the standard deviation factor of the population.
[0051] The number of random individuals introduced is determined by the standard deviation factor of the population: when the standard deviation indicates insufficient population diversity, the number of random individuals is appropriately increased to promote exploration; when the population already has high diversity, the number of random individuals introduced is reduced. Furthermore, the number of newly generated individuals also needs to be set within a reasonable range.
[0052] Therefore, the following equation is designed to calculate the number of randomly generated individuals: Where ε is the set threshold, σ mean is the standard deviation factor of the population, and k is the proportionality coefficient.
[0053] Specifically, in one embodiment of the present invention, based on σ meanThe pseudocode for the new individual generation strategy is shown in Algorithm 3:
[0054] Specifically, the steps for generating a new population by selecting a certain number of individuals from the merged population according to the elite selection strategy are as follows: Randomly generate numoff individuals in the search space, i.e.: x i =L R +(U R -L R )×random.
[0055] Among them, L R and U R These are the upper and lower bounds of the combination parameters, respectively. `random` is a random number between 0 and 1. As can be seen from the above method for generating new individuals, the newly added individuals are randomly distributed in the search space, which results in good diversity in the initial population.
[0056] In one embodiment of the present invention, the method further includes step S500: constructing an evaluation function using a penalty function to obtain the lowest reflection loss within a specific frequency range, wherein the penalty function penalizes solutions that violate constraints by adding an extra term, wherein the penalty function is:
[0057] Wherein, the function f is an approximate value of the reflectivity RL obtained by predicting through a preset neural network model, d, p, a, w, r are the combination parameters of the absorbing material, where d represents the thickness of the medium, p represents the pattern arrangement period, a represents the outer dimension of the square ring, w represents the ring width of the square ring, r represents the surface resistance of the square ring, and fq i Indicates frequency range, dB i θ is the reflectivity threshold at the i-th frequency, and θ is the penalty coefficient.
[0058] Although the present invention has been described in detail through the preferred embodiments above, it should be understood that the above description should not be considered as a limitation of the present invention. Various modifications and substitutions to the present invention will be apparent to those skilled in the art after reading the above description. Therefore, the scope of protection of the present invention should be defined by the appended claims.
[0059] The present invention will be further illustrated below with reference to specific application scenarios. It should be understood that the following embodiments are only for illustrating the present invention and are not intended to limit the present invention.
[0060] For example, the present invention trains a BP neural network model as follows: Figure 4As shown: The model has 11 input layer parameters, including d1, d2, d3, p, a1, a2, a3, w, r1, r2, and r3, totaling 11 parameters. Here, d1, d2, and d3 represent the thicknesses of the first, second, and third media layers, respectively; d1 + d2 + d3 represents the total thickness of the three media layers; a1, a2, and a3 represent the outer dimensions of the first, second, and third square rings, respectively; and r1, r2, and r3 represent the surface resistance values of the first, second, and third square rings, respectively. The output layer is the reflectivity RL variation at frequencies from 2 to 18 GHz. The initial learning rate was 0.0001, epochs = 50, and the loss function was MSE. The Adam optimizer and BatchNormalization were used for standardization. A total of 60,000 data points were obtained through dataset preparation. The entire dataset was divided into 5 parts using K-fold cross-validation, and validation was performed every 100 epochs. The optimal model parameters were saved based on the validation loss. Finally, the evaluation was performed on the test set.
[0061] It should be noted that those skilled in the art can also refer to the parameters of the BP neural network model provided in this invention and choose to use similar, close, or specially designed parameters, which will still fall within the scope of protection of this invention. The structural parameter settings of the BP neural network are as follows:
[0062] As shown in the table: Table 1
[0063] To verify the performance of the proposed algorithm, it was tested on this microwave absorbing material optimization design problem. To obtain reliable optimization results, 30 rounds of experiments were conducted for each problem, with the maximum number of iterations T per round for each algorithm set to 100. Comparisons were made with traditional GA (General Genetic Algorithm), PSO (Particle Swarm Optimization), DE (Differential Evolutionary Algorithm), and the latest AWPSO (Adaptive Weighted Particle Swarm Optimization) algorithm to find the parameter combination of the absorber with the lowest reflectivity in the 2-18 GHz frequency range. The Our-GA algorithm employed two-point crossover and tournament selection, with crossover probability (Pc) and mutation probability (Pm) set to 0.8 and 0, respectively. The PSO algorithm had an acceleration factor of 2.02 and used adaptive inertial weights, while the DE algorithm had a crossover probability of 0.8 and a scaling factor F of 0.5. All algorithms compared in this section used the same population size: M = 40. They also used the same termination condition.
[0064] To better illustrate the structural design method of the present invention, specific embodiments 1, 2 and 3 are provided below.
[0065] In Example 1, the objective function is generated by inputting the target reflectivity within a specific frequency band, along with constraints on other structural variables:
[0066] To meet the diverse design needs, such as Figure 5 and Figure 6 As shown, to achieve a reflectivity RL less than -8GHz and -10GHz in the 2-18GHz range, θ was set to 200 and 100 respectively. The optimal values after 30 iterations were analyzed, as shown in Tables 2 and 3. Table 2 represents the penalty value db. i Table 3 shows the penalty value in dB for the -8GHz optimization results. i Results optimized for -10GHz.
[0067] Table 2
[0068] Table 3
[0069] The predicted values of each algorithm are shown in Table 4. The Our-GA algorithm meets the design requirement that the reflectivity RL is less than that of -8GHz and -10GHz in the 2-18GHz range. To verify the prediction quality of the BP network, finite element simulations were performed on the predicted values of the solution obtained by the Our-GA algorithm at -8GHz. The MAPE between the two was 12.9%.
[0070] Table 4
[0071] As can be seen from Tables 2 and 3, the DE and Our-GA algorithms have better Average values. Figure 7 and Figure 8 Further comparison of the distribution of DE and Our-GA solutions reveals that, under the premise of reflectivity RL being less than -8GHz and -10GHz, the Our-GA solution exhibits greater diversity, generating more design options and better meeting the requirements of Task 1.
[0072] In Example 2, given the target reflectivity within a specific frequency band, the minimum value of d1+d2+d3 (total thickness) is provided. The objective function provided in the above examples... Based on this, a penalty function is added for d1+d2+d3, that is:
[0073] Where θ1 and θ2 are set to 100 and 20 respectively, and beta is the adjustment parameter for each frequency band. Tables 5 and 6 show the penalty value db. iThe optimization results are shown in Table 5, where the penalty value db is represented. i Table 6 shows the penalty value in dB for the -8GHz optimization results. i Results optimized for -10GHz.
[0074] Table 5
[0075] Table 6
[0076] As can be seen from Tables 5 and 6, the DE30 solution has the smallest Average, and the Our-GA iteration has the smallest Best, which satisfies the condition in the 2-18 GHz range (db). i Under the requirement of less than -8GHz, d sum The minimum value.
[0077] See Tables 7 and 8 for further details. Table 7 is a database. i Minimum statistics for frequencies below -8 GHz, Table 7 shows the data in dB. i Minimum statistical data below -10GHz.
[0078] Table 7
[0079] Table 8
[0080] As can be seen, in db i Below -8GHz, the d-sumBest values of various algorithms are basically the same, db i At frequencies below -10 GHz, Our-GA has the lowest d-sum Best and d-sum Average.
[0081] In Example 3, the input is the total thickness value d within a specific frequency band. sum The objective function provided in the above embodiment gives the minimum reflectivity under these constraints. Adding to d based on sum The limitation is: Subject to: d1+d2+d3=d sum
[0082] Among them, when db i At -8GHz and -10GHz, d sum Limited to 15mm and 20mm, with θ set to 200 and 100, the optimization results are shown in Tables 9 and 10. Figure 9 and Figure 10As shown, Table 9 contains the db data. i Minimum statistics for frequencies below -8 GHz, Table 10 shows the data in dB. i Minimum statistical data below -10GHz.
[0083] Table 9
[0084] Table 10
[0085] From Tables 9 and 10, we can conclude that Our-GA in db i Best is minimum at -8GHz, dB i The Average value is minimum at -10GHz; by Figure 9 and Figure 10 It can be concluded that Our-GA has the lowest average fitness.
[0086] Combining Examples 1, 2, and 3, it can be concluded that the Our-GA algorithm provided by this invention excels in solution diversity, finding the minimum RL value, and d sum The overall solution under constraints exhibits excellent quality and meets the design requirements.
[0087] This invention presents a simple and easy-to-implement design method for microwave absorbing metamaterials using artificial neural networks and improved genetic algorithms. This method shortens the fabrication cycle, significantly reduces computational costs, and enables global search within a multivariate parameter space to optimize structural parameters, quickly and accurately finding the optimal design scheme at the operating frequency. Finally, under the same optimization parameter settings, the neural network output is compared with the simulation results. The mean absolute percentage error (MAPE) between the two is 12.9%, verifying the efficiency and accuracy of the optimization design. This method significantly reduces experimental costs and accelerates the design process of microwave absorbing metamaterials. The MAPE is calculated by averaging the percentage deviation between the predicted value and the simulation software value at each frequency.
[0088] like Figure 11 As shown, a second aspect of the present invention provides a forward design system for microwave absorbing metamaterials, comprising: An initial sample acquisition unit is used at least to uniformly sample and acquire an initial sample within the geometric space of the sample, and to simulate the initial sample at a preset frequency using preset simulation software to obtain the corresponding initial sample absorption performance parameters. The high error sample acquisition unit is used to randomly sample and acquire high error samples in the geometric space of the sample, and to simulate the high error samples at a preset frequency using preset simulation software to obtain the high error sample absorption performance parameters of the corresponding high error samples. A high-quality sample acquisition unit is used to at least increase the geometric space constraints of the sample, acquire high-quality samples, and simulate the high-quality samples at a preset frequency using preset simulation software to obtain the high-quality sample absorption performance parameters of the corresponding high-quality samples. The model training and tuning unit is used at least to fill the initial sample absorption performance parameters into the total dataset to train a preset neural network model; the high-error sample absorption performance parameters and the high-quality sample absorption performance parameters are respectively filled into the total dataset to gradually adjust the accuracy of the trained neural network model.
[0089] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.
[0090] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of each example have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.
[0091] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for forward design of a wave-absorbing metamaterial, characterized in that, include: Initial samples are obtained by uniformly sampling within the geometric space of the sample. The initial samples are then simulated at a preset frequency using preset simulation software to obtain the corresponding initial sample absorption performance parameters. These initial sample absorption performance parameters are then filled into the total dataset to train a preset neural network model. High-error samples are randomly sampled within the geometric space of the sample. The high-error samples are then simulated at a preset frequency using preset simulation software to obtain the high-error sample absorption performance parameters. Increase the constraints of the geometric space of the sample to obtain high-quality samples. Simulate the high-quality samples at a preset frequency using preset simulation software to obtain the high-quality sample absorption performance parameters of the corresponding high-quality samples. The absorption performance parameters of the high-error samples and the absorption performance parameters of the high-quality samples are respectively filled into the total dataset to gradually adjust the accuracy of the neural network model.
2. The method of claim 1, wherein, The preset neural network model is a backpropagation neural network model.
3. The method of claim 1, wherein, The process of uniformly sampling within the geometric space of the sample to obtain initial samples, simulating the initial samples at a preset frequency using preset simulation software to obtain corresponding initial sample absorption performance parameters, and then filling the initial sample absorption performance parameters into the total dataset to train a preset neural network model includes: Initial samples are generated by uniformly sampling within the geometric space of the samples using the Latin hypercube method. The corresponding absorption performance parameters are obtained by simulating at a preset frequency using electromagnetic field simulation software. The initial sample absorption performance parameters are filled into the total dataset to train the backpropagation neural network model.
4. The method of claim 1, wherein, The step of randomly sampling within the geometric space of the sample to obtain high-error samples includes: Screening out the individual x with the largest error from the simulation value in the sample, and selecting k nearest neighbors of the individual x nn ; By x new = x + δ x (x - x nn ) in the preset range of the individual region with the largest prediction error of the preset neural network model to generate a new sample x new as a high error sample.
5. The method of claim 1, wherein, The method of increasing the geometric space constraints of the samples to obtain high-quality samples includes: Using a genetic algorithm, the original population and corresponding individuals are configured. The binary string of each individual is divided into several parts according to the number of parameters, and each part represents a sample absorption performance parameter. Neighbor sampling was performed on the original population to fill in auxiliary populations; The elite selection strategy is used to select samples in the auxiliary population with the same number as the original population. Set the crossover probability and randomly generate the random probability; The crossover probability and random probability are compared, and parents are selectively obtained from the original population and auxiliary population to generate offspring populations, which are then used as high-quality samples.
6. The method of forward design of a wave-absorbing metamaterial according to claim 5, wherein, The step of comparing crossover probabilities and random probabilities, selectively obtaining parent generations from the original population and auxiliary populations to generate offspring populations, and using the offspring populations as high-quality samples includes: If the crossover probability is greater than the random probability, two individuals are randomly selected from the original population as parents. The two points intersect and fill the offspring population; If the crossover probability is not greater than the random probability, select one individual from the original population and one individual from the auxiliary population as the parent generation; The two points intersect and fill the offspring population.
7. The metamaterial absorber forward design method of claim 6, wherein, It also includes obtaining and determining the number of random individuals in the offspring population based on the standard deviation factor of the population.
8. The method of forward design of a wave-absorbing metamaterial according to claim 1, wherein, It also includes using a penalty function to construct an evaluation function to obtain the lowest reflection loss within a specific frequency range.
9. The metamaterial absorber forward design method of claim 8, wherein, In the penalty function, solutions that violate the constraints are penalized by adding an extra term, wherein the penalty function is: The function f is an approximation of the reflectivity RL predicted by a preset neural network model, d, p, a, w, and r are the combination parameters of the wave-absorbing material, and fq i represents a frequency range, db i is the reflectivity threshold of the i-th frequency, and θ is a penalty coefficient.
10. A metamaterial absorber forward design system, comprising: include: A high-error sample acquisition unit is used at least to uniformly sample and acquire an initial sample in the geometric space of the sample, and to simulate the initial sample at a preset frequency using preset simulation software to obtain the corresponding initial sample absorption performance parameters. The high error sample acquisition unit is used to randomly sample and acquire high error samples in the geometric space of the sample, and to simulate the high error samples at a preset frequency using preset simulation software to obtain the high error sample absorption performance parameters of the corresponding high error samples. A high-quality sample acquisition unit is used to at least increase the geometric space constraints of the sample, acquire high-quality samples, and simulate the high-quality samples at a preset frequency using preset simulation software to obtain the high-quality sample absorption performance parameters of the corresponding high-quality samples. The model training and tuning unit is used at least to fill the initial sample absorption performance parameters into the total dataset to train a preset neural network model. The high-error samples fill the absorption performance parameters and the high-quality sample absorption performance parameters into the total dataset respectively, so as to gradually adjust the accuracy of the trained neural network model.
