Frequency selection structure design method based on scaling data and machine learning
By combining the principle of physical scaling with machine learning, an initial dataset is scientifically generated, solving the problems of high data acquisition costs and low efficiency in frequency selection structure design. This enables fast and accurate frequency selection structure design, improving design efficiency and success rate, and reducing costs and technical barriers.
Patent Information
- Application Number
- CN202511997087.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-27
- Publication Date
- 2026-05-15
AI Technical Summary
Existing frequency selection structure design methods suffer from high costs and low efficiency in the early stages of design, and rely heavily on the subjective experience of designers, resulting in long design cycles, high resource consumption, and difficulty in guaranteeing global optimality.
Combining the principles of physical scaling and machine learning, an initial dataset is scientifically generated, initial parameters are determined using the scaling principle, the range of parameter variation is set in conjunction with engineering constraints, a sample dataset is constructed using an efficient sampling method, an initial surrogate model is trained, and global optimization is performed using an intelligent optimization algorithm, ultimately verifying the optimal structural parameters.
Significantly improves design efficiency, reduces the number of simulation samples by 80%-90%, shortens the design cycle from several weeks to several days, reduces R&D costs, increases the design success rate, lowers the technical threshold, and enables engineers with little experience to quickly design high-performance frequency-selective structures.
Smart Images

Figure CN122046568A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of electromagnetic metamaterials and microwave devices, specifically relating to a frequency-selective structure design method based on scaled-down data and machine learning. Background Technology
[0002] Frequency selective structures (frequency selective surfaces) are two-dimensional periodic arrays capable of filtering electromagnetic waves, playing a crucial role in fields such as radomes, stealth technology, and satellite communications. Their performance is determined by a combination of factors, including element geometry, period, and dielectric material; the design process is essentially a complex multidimensional parameter optimization problem.
[0003] Traditional design methods mainly rely on the designer's experience and parameter scanning method, that is, to traverse the design space through a large number of repeated full-wave electromagnetic simulations. This method has the following drawbacks: (1) It is very time-consuming: a full-wave simulation usually takes several minutes to several hours, and it may take several weeks to complete a design; (2) It is computationally expensive: it requires high-performance computing resources and expensive commercial simulation software; (3) It is easy to get trapped in local optima: in a multi-parameter, strongly coupled design space, it is difficult to guarantee that the global optimal solution can be found.
[0004] To address these issues, machine learning-aided design methods have emerged. The core idea is to first generate a batch of data through simulation, and then use this data to train a surrogate model to replace time-consuming full-wave simulations. However, existing machine learning methods generally suffer from a key bottleneck: the construction of the initial dataset. Training an accurate surrogate model requires a large and evenly distributed dataset. Currently, obtaining this dataset often involves grid sampling or random sampling, which still requires hundreds or thousands of simulations and does not fundamentally solve the cost problem of data acquisition, the so-called "cold start" problem. Designers often need to rely on experience to set a wide range of parameters for sampling, which is not only inefficient, but also results in a large portion of the data being "poor-quality data" useless for final optimization, impacting the efficiency of the entire design process.
[0005] Existing design methods for frequency selection structures, whether traditional parametric scanning methods or emerging machine learning-assisted methods, all face the problems of high costs and low efficiency in early-stage data acquisition, and heavy reliance on the designer's subjective experience to determine the initial exploration scope. This results in a long design cycle, high resource consumption, and difficulty in guaranteeing the global optimality of the design results. Summary of the Invention
[0006] This invention addresses the problems of high cost, low efficiency, and heavy reliance on the designer's subjective experience in determining the initial exploration scope during the early stages of frequency selection structure design. It provides an efficient, scientific, and low-cost method for frequency selection structure design. By introducing the principle of physical scaling as prior knowledge, it scientifically guides the generation of the initial dataset, aiming to overcome the "cold start" problem in machine learning-assisted design. It constructs a high-quality initial dataset with minimal simulation costs, thereby significantly improving the training and optimization efficiency of subsequent surrogate models, and ultimately achieving rapid and accurate design of frequency selection structures.
[0007] The technical solution of this invention is implemented as follows: A frequency selection structure design method based on scaled-down data and machine learning includes the following steps: Based on existing structural parameters, the initial parameters of the frequency-selective structure are determined using the scaling principle. Based on engineering constraints, set the range of variation for frequency-selective structural parameters; Collect sample data based on the above parameter variation range and construct a sample dataset; Construct a machine learning algorithm model and train the initial agent model using the above sample dataset; By combining the initial agent model, an intelligent optimization algorithm is used to globally optimize the frequency-selective structural parameters and obtain the optimal structural parameters. Verify whether the above optimal structural parameters meet the design objectives. If they do, end the process. If the requirements are not met, the current optimal structural parameters are placed in the sample set, the surrogate model is retrained, and the above process is repeated until the design goal is met.
[0008] As a further aspect of the present invention: based on existing structural parameters, the initial parameters of the frequency-selective structure are determined using the scaling principle, specifically as follows: Based on existing structural parameters, the initial structural parameters of the target are determined by using the scaling principle to achieve frequency band shifting.
[0009] As a further aspect of the present invention: A frequency-selective surface reference design with known performance is obtained, operating at a frequency... f 1. The dielectric constant of the substrate is ε r1 According to the target of the new design: frequency f 2. Dielectric constant ε r2 Through formula The center point of the initial structural geometric parameters of the newly designed structure was calculated. L 2, of which, L 1 represents the current structural geometry parameter.
[0010] As a further aspect of the present invention: In conjunction with engineering constraints, the variation range of the frequency-selective structural parameters is set, specifically as follows: Using the frequency calculated based on the scaling principle as a reference, a preset perturbation range of less than ±20% is applied to each geometric parameter to construct a multidimensional, bounded initial parameter sampling space.
[0011] As a further aspect of the present invention: Sample data is collected based on the aforementioned parameter variation range to construct a sample dataset, specifically as follows: Latin hypercube sampling is used to collect samples from high-potential spaces, and training samples are obtained through full-wave simulation to construct a sample dataset. Each training sample includes a combination of structural parameters and a combination of performance parameters.
[0012] As a further aspect of this invention: a machine learning algorithm model is constructed, and an initial agent model is trained using the aforementioned sample dataset, specifically as follows: The combination of structural parameters generated by sampling is used as the model input, and the corresponding combination of performance parameters is used as the model output. Gaussian process regression, neural network or support vector regression algorithms are selected for initial surrogate model training.
[0013] As a further aspect of this invention: combining the initial proxy model, a smart optimization algorithm is used to perform global optimization of the frequency-selective structure parameters to obtain the optimal structure parameters, specifically: An optimization objective function is constructed, and combined with the initial surrogate model, a genetic algorithm, particle swarm optimization algorithm, and differential evolution algorithm are used to quickly iterate and search on the basis of the initial surrogate model to determine the optimal combination of structural parameters that satisfies the optimization objective.
[0014] As a further aspect of the present invention: verifying whether the above-mentioned optimal structural parameters meet the design objectives, specifically: Substitute the above optimal combination of structural parameters into the full-wave simulation tool to verify whether the electromagnetic simulation results meet the design objectives.
[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention combines the principle of physical scaling with machine learning experimental design theory to guide the entire process of initial data acquisition in frequency-selective structure design. Its core protection scope covers: a framework that uses the principle of physical scaling as prior knowledge to "start" the machine learning design process; a complete and systematic method from "obtaining a reference design" to "applying the scaling formula to calculate the center point," and then to "constructing the perturbation space" and using "efficient sampling methods (such as LHS)" to generate the dataset; and the application of this method in the design of various frequency-selective structures such as frequency-selective surfaces, metamaterial absorbers, and miniaturized antennas.
[0016] 2. This invention brings significant benefits to the design of frequency-selective structures through a scientific and efficient new paradigm for data acquisition: Significantly improve design efficiency: By accurately sampling the "potential optimal solution region" located by the scaling principle, the number of simulation samples required is reduced by 80%-90% compared to traditional methods, shortening the design cycle from several weeks to several days.
[0017] Significantly reduced R&D costs: The sharp reduction in the number of simulations directly saves expensive computing resources and the cost of using commercial software.
[0018] Improving design success rate: High-quality datasets help train more accurate surrogate models, making it easier to find the globally optimal solution that meets performance metrics and avoid design failures caused by getting stuck in local optima in traditional methods.
[0019] Lowering the technical threshold: This method is procedural and systematic, reducing the reliance on the subjective experience of design engineers, enabling engineers with less experience to quickly and scientifically carry out the design of high-performance frequency-selective structures.
[0020] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0021] Figure 1 This is a schematic diagram of the frequency selection structure design algorithm of a preferred embodiment of the present invention; Figure 2 This is a schematic diagram of the overall structure of the frequency selection structure according to a preferred embodiment of the present invention; Figure 3 This is a schematic diagram of the first metal layer of the frequency-selective structure according to a preferred embodiment of the present invention; Figure 4 This is a schematic diagram of the second metal layer of the frequency-selective structure according to a preferred embodiment of the present invention; Figure 5 This is a schematic diagram of the third metal layer of the frequency-selective structure in a preferred embodiment of the present invention; Figure 6 This is a schematic diagram of the S-parameters of the frequency-selective structure simulation according to a preferred embodiment of the present invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be described in more detail below with reference to the accompanying drawings.
[0023] In the accompanying drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, of the embodiments of the present invention.
[0024] The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.
[0025] The following is in conjunction with the appendix Figure 1-6 The embodiments of the present invention will be described in detail below.
[0026] Example 1 This invention provides a frequency selection structure design method based on scaled data and machine learning, which includes the following steps: Based on existing structural parameters, the initial parameters of the frequency-selective structure are determined using the scaling principle. Based on engineering constraints, set the range of variation for frequency-selective structural parameters; Collect sample data based on the above parameter variation range and construct a sample dataset; Construct a machine learning algorithm model and train the initial agent model using the above sample dataset; By combining the initial agent model, an intelligent optimization algorithm is used to globally optimize the frequency-selective structural parameters and obtain the optimal structural parameters. Verify whether the above optimal structural parameters meet the design objectives. If they do, end the process. If the requirements are not met, the current optimal structural parameters are placed in the sample set, the surrogate model is retrained, and the above process is repeated until the design goal is met.
[0027] Furthermore, based on the existing structural parameters, the initial parameters of the frequency-selective structure are determined using the scaling principle, specifically: Based on existing structural parameters, the initial structural parameters of the target are determined by using the scaling principle to achieve frequency band shifting.
[0028] Furthermore, a frequency-selective surface reference design with known performance is obtained, operating at a frequency... f 1. The dielectric constant of the substrate is ε r1 According to the target of the new design: frequency f 2. Dielectric constant ε r2 Through formula The center point of the initial structural geometric parameters of the newly designed structure was calculated. L 2, of which, L 1 represents the current structural geometry parameter.
[0029] Furthermore, considering engineering constraints, the range of variation for frequency-selective structural parameters is set as follows: Using the frequency calculated based on the scaling principle as a reference, a preset perturbation range of less than ±20% is applied to each geometric parameter to construct a multidimensional, bounded initial parameter sampling space.
[0030] Furthermore, based on the aforementioned parameter variation range, sample data is collected to construct a sample dataset, specifically as follows: Latin hypercube sampling is used to collect samples from high-potential spaces, and training samples are obtained through full-wave simulation to construct a sample dataset. Each training sample includes a combination of structural parameters and a combination of performance parameters.
[0031] Furthermore, a machine learning algorithm model is constructed, and the initial agent model is trained using the aforementioned sample dataset, specifically as follows: The combination of structural parameters generated by sampling is used as the model input, and the corresponding combination of performance parameters is used as the model output. Gaussian process regression, neural network or support vector regression algorithms are selected for initial surrogate model training.
[0032] Furthermore, combining the initial proxy model, an intelligent optimization algorithm is used to globally optimize the frequency-selective structure parameters to obtain the optimal structure parameters, specifically: An optimization objective function is constructed, and combined with the initial surrogate model, a genetic algorithm, particle swarm optimization algorithm, and differential evolution algorithm are used to quickly iterate and search on the basis of the initial surrogate model to determine the optimal combination of structural parameters that satisfies the optimization objective.
[0033] Furthermore, we verify whether the above optimal structural parameters meet the design objectives, specifically as follows: Substitute the above optimal combination of structural parameters into the full-wave simulation tool to verify whether the electromagnetic simulation results meet the design objectives.
[0034] This invention relates to a comprehensive method that combines physical scaling principles with machine learning experimental design theory to guide the initial data acquisition process in frequency-selective structure design. Its core scope covers: a framework that uses physical scaling principles as prior knowledge to "initiate" the machine learning design process; a complete and systematic method from "obtaining a reference design" to "applying scaling formulas to calculate the center point," and then to "constructing the perturbation space" and using "efficient sampling methods (such as LHS)" to generate the dataset; and the application of this method in the design of various frequency-selective structures such as frequency-selective surfaces, metamaterial absorbers, and miniaturized antennas.
[0035] Example 2 This invention provides a frequency-selective surface design method based on scaled data and machine learning. The method includes the following steps: generating an initial center point using physical scaling principles; constructing a high-potential sampling space around the center point; conducting efficient experimental design sampling within the space; constructing an initial dataset through simulation; and training a machine learning surrogate model using this dataset for rapid optimization design. This method organically combines physical laws with data science, solving the problems of blind and inefficient data acquisition in traditional design methods. Based on existing structural parameters, the method uses physical scaling principles to achieve frequency band shifting to determine the target initial structural center point. A small-range parameter variation space is set around the center point in conjunction with engineering constraints. Latin hypercube sampling is used to collect samples from the high-potential space, and training samples are obtained through full-wave simulation. A Gaussian process regression algorithm is used to establish a training surrogate model to replace the full-wave simulation, based on the collected samples. On this basis, an intelligent optimization algorithm is used for global optimization. The current optimal solution is substituted into the full-wave simulation tool to verify whether the electromagnetic simulation results meet the design objectives. If they do, the process ends; if not, the current optimal sample is placed in the sample set, the surrogate model is retrained, and the above process is repeated until the design objectives are met.
[0036] Secondary inventive point: The specific steps for generating the initial center point using the physical scaling principle are as follows: Obtain a frequency-selective surface reference design with known performance (operating at a frequency of...). f 1. The dielectric constant of the substrate is ε r1 According to the goals of the new design (frequency) f 2. Dielectric constant ε r2 ), through formula The initial geometric parameters center point of the new design were calculated. L 2.
[0037] The three secondary invention points are as follows: The steps for constructing a high-potential sampling space are as follows: Using the frequency calculated by the scaling principle as a reference, the center point of the initial geometric parameters of the surface is selected, and a preset perturbation range of less than ±20% is applied to each geometric parameter, thereby constructing a multi-dimensional, bounded initial parameter sampling space.
[0038] 4. Secondary Invention Points: The efficient experimental design sampling method is the Latin Hypercube Sampling (LHS) method, which can maximize the coverage of the sampling space with the fewest sample points while ensuring that the samples are uniformly distributed in all dimensions.
[0039] Five key invention points: The steps for training a machine learning proxy model using a dataset include: using the sampled parameter combination as the model input, using the corresponding full-wave simulation results as the model output, and using algorithms such as Gaussian process regression (GPR), neural network (ANN), or support vector regression (SVR) for model training.
[0040] 6 Secondary Invention Points: The design method described in the 1 core invention point is characterized in that: the rapid optimization design step specifically involves: using the trained machine learning surrogate model as a fitness function or objective function, and combining one or more global optimization algorithms (such as genetic algorithm, particle swarm optimization algorithm, differential evolution algorithm, etc.) to perform rapid iterative search on the surrogate model to determine the optimal combination of structural parameters that meets the preset performance indicators.
[0041] Example 3 The algorithm implementation process is as follows: Figure 1 As shown. Using... Figure 2 The algorithm flow of this invention is verified using a three-layer frequency selective surface structure. This frequency selective surface comprises three conductor layers, designated as a first conductor layer 10, a second conductor layer 20, and a third conductor layer 30. A first dielectric layer 12 is disposed between the first conductor layer 10 and the second conductor layer 20, and a second dielectric layer 23 is disposed between the second conductor layer 20 and the third conductor layer 30. The first conductor layer 10 includes a hexagonal patch structure 101. The second conductor layer 20 has an inner hexagonal ring 201 and an outer hexagonal ring 202, with the inner hexagonal ring 201 disposed inside the outer hexagonal ring 202, and the two are equally spaced. The third conductor layer includes a hexagonal patch structure 301. This example aims to quickly migrate a reference frequency selective surface design operating at 29 GHz to a new frequency selective surface operating at 22.9 GHz, primarily used for communication and data links.
[0042] according to Figure 1 The process shown first follows the scaling principle formula of secondary invention point 2. Calculate the structural parameters corresponding to 22.9 GHz under the new dielectric substrate. p , p_patch , w_loop 1, w_loop 2, p_loop , h After obtaining the calculation results, based on these initial calculation results, expand the secondary invention point 3 by less than ±20% on both sides, for example, 0.5 mm. If the structural parameter is less than 0.5 mm, for example... w_loopIf 1 = 0.2 mm, then the lower limit of its variation is set to 0.1 mm, and the upper limit to 0.7 mm. Based on secondary invention point 4, the LHS method is used to generate 50 initial samples in this space. Typically, the number of initial samples is between 5 and 10 times the number of structural parameters. Machine learning algorithms, such as Gaussian process regression, are used to train a surrogate model. This model replaces full-wave simulation. Heuristic intelligent optimization algorithms, such as genetic algorithms and particle swarm optimization, are used to achieve iterative optimization. If the design goal is not met, the current optimization result is added to the dataset, the surrogate model is retrained, and the optimization algorithm is executed again to determine if the goal is met, until the set optimization goal is achieved. Figure 1 The algorithm flow shown is for... Figures 2-5 The frequency selection structure shown is optimized to finally obtain Figure 6 The optimization results are shown.
[0043] Thus, the objective of this invention has been achieved.
[0044] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A frequency-selective structure design method based on scaled-down data and machine learning, characterized in that, Includes the following steps: Based on existing structural parameters, the initial parameters of the frequency-selective structure are determined using the scaling principle. Based on engineering constraints, set the range of variation for frequency-selective structural parameters; Collect sample data based on the above parameter variation range and construct a sample dataset; Construct a machine learning algorithm model and train the initial agent model using the above sample dataset; By combining the initial agent model, an intelligent optimization algorithm is used to globally optimize the frequency-selective structural parameters and obtain the optimal structural parameters. Verify whether the above optimal structural parameters meet the design objectives. If they do, end the process. If the requirements are not met, the current optimal structural parameters are placed in the sample set, the surrogate model is retrained, and the above process is repeated until the design goal is met.
2. The frequency selection structure design method based on scaled data and machine learning according to claim 1, characterized in that, Based on existing structural parameters, the initial parameters of the frequency-selective structure are determined using the scaling principle, specifically: Based on existing structural parameters, the initial structural parameters of the target are determined by using the scaling principle to achieve frequency band shifting.
3. The frequency selection structure design method based on scaled data and machine learning according to claim 2, characterized in that, Obtain a frequency-selective surface reference design with known performance, operating at a frequency f 1. The dielectric constant of the substrate is ε r1 According to the target of the new design: frequency f 2. Dielectric constant ε r2 Through formula The center point of the initial structural geometric parameters of the newly designed structure was calculated. L 2, of which, L 1 represents the current structural geometry parameter.
4. The frequency selection structure design method based on scaled data and machine learning according to claim 3, characterized in that, Based on engineering constraints, the range of variation for frequency-selective structural parameters is set as follows: Using the frequency calculated based on the scaling principle as a reference, a preset perturbation range of less than ±20% is applied to each geometric parameter to construct a multidimensional, bounded initial parameter sampling space.
5. The frequency selection structure design method based on scaled data and machine learning according to claim 1, characterized in that, Based on the above parameter variation range, sample data was collected to construct a sample dataset, specifically as follows: Latin hypercube sampling is used to collect samples from high-potential spaces, and training samples are obtained through full-wave simulation to construct a sample dataset. Each training sample includes a combination of structural parameters and a combination of performance parameters.
6. The frequency selection structure design method based on scaled data and machine learning according to claim 5, characterized in that, To construct a machine learning algorithm model, and using the aforementioned sample dataset, train the initial agent model as follows: The combination of structural parameters generated by sampling is used as the model input, and the corresponding combination of performance parameters is used as the model output. Gaussian process regression, neural network or support vector regression algorithms are selected for initial surrogate model training.
7. The frequency selection structure design method based on scaled data and machine learning according to claim 6, characterized in that, Combining the initial proxy model, an intelligent optimization algorithm is used to globally optimize the frequency-selective structure parameters to obtain the optimal structure parameters, specifically: An optimization objective function is constructed, and combined with the initial surrogate model, a genetic algorithm, particle swarm optimization algorithm, and differential evolution algorithm are used to quickly iterate and search on the basis of the initial surrogate model to determine the optimal combination of structural parameters that satisfies the optimization objective.
8. The frequency selection structure design method based on scaled data and machine learning according to claim 7, characterized in that, To verify whether the above optimal structural parameters meet the design objectives, the following steps are taken: Substitute the above optimal combination of structural parameters into the full-wave simulation tool to verify whether the electromagnetic simulation results meet the design objectives.
