Wave-absorbing material design method based on neural network guidance

Abstract

The invention discloses a wave-absorbing material design method based on neural network guidance, and belongs to the field of intelligent material design. The method comprises the following steps: selecting actually measured magnetic conductivity frequency dispersion data of a magnetic material as a boundary condition, constructing a parameter space containing magnetic conductivity, dielectric constant, material thickness and frequency, and calculating a reflection loss RL value through a transmission line model; gPU acceleration tensor operation is adopted to generate a three-dimensional parameter space, and discretization processing is carried out on the parameter space; implementing a double-task screening strategy to extract an optimal dielectric constant combination and an effective region boundary; constructing a neural network model with physical constraints, and inputting thickness and frequency prediction dielectric constant parameters; and generating solution set distribution to guide material design through inversion verification. A physical model and deep learning are fused, an Einstein summation method and a broadcast mechanism are adopted to realize hundred million-level parameter parallel calculation, the problem of multi-parameter coupling optimization is solved, the physical rationality of output parameters is ensured through a softplus activation function, and the design efficiency and precision of the magnetic wave-absorbing material are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent material design, and provides a method for designing microwave absorbing materials guided by neural networks. Background Art

[0002] In recent years, with the continuous upgrading of radar monitoring systems, stealth technology, as a powerful anti-reconnaissance method, is of great significance for maintaining the security of weapon systems. Developing light, thin, and highly efficient microwave absorbing coatings has become a current research hotspot. However, most current dielectric materials are difficult to meet the requirements of being light, thin, and broadband, while magnetic materials can effectively adjust the impedance matching characteristics of materials by introducing magnetic permeability, thereby achieving broadband and efficient absorption. Nevertheless, there is still a lack of systematic theoretical guidance on how to precisely customize the design of magnetic microwave absorbing materials.

[0003] Although the transmission line model can correlate the electromagnetic parameters of microwave absorbing materials with their microwave absorption performance, multi-parameter coupling results in an extremely complex mapping relationship. Previously, our team established a "dielectric constant genome" of non-magnetic materials through numerical screening and fitting, successfully guiding the design of dielectric microwave absorbing materials (Patent Nos.: ZL 202310167584.0 and ZL202310227024.X). However, for magnetic materials, the introduction of the relative complex magnetic permeability with dispersion characteristics makes the complexity of this method increase exponentially and become difficult to implement.

[0004] The magnetic permeability of existing magnetic microwave absorbing materials shows a similar change trend and small fluctuations in the range of 1 - 18 GHz, but the coupling optimization with dielectric parameters still faces challenges. Traditional electromagnetic parameter synchronous optimization methods are difficult to handle the complex non-linear relationship between magnetic permeability and dielectric constant. In addition, the parameter space discretization calculation based on the transmission line model involves a huge number of combinations (such as 230.4 million), and the traditional nested loop calculation is inefficient. Although the calculation speed can be significantly improved by GPU acceleration, the generated dielectric constant data set shows complex non-linear characteristics, and traditional function fitting methods cannot effectively guide material design due to the lack of a simple mathematical form.

[0005] In summary, there is an urgent need for a new method that combines physical mechanisms and data-driven approaches to break through the bottleneck of electromagnetic parameter decoupling optimization and intelligent design of magnetic microwave absorbing materials. Summary of the Invention

[0006] The purpose of the present invention is to solve the optimization problem in the design of magnetic microwave absorbing materials caused by multi-parameter coupling of magnetic permeability and dielectric constant.

[0007] To achieve the above purpose, the present invention adopts the following technical solutions:

[0008] A method for designing microwave absorbing materials guided by neural networks, comprising the following steps:

[0009] Step (1): Based on the magnetic permeability dispersion characteristics of magnetic wave-absorbing materials, select the measured magnetic permeability of magnetic materials with dispersion characteristics as the reference boundary condition;

[0010] Step (2): Construct a high-throughput screening parameter space including magnetic permeability μ r , dielectric constant ε r , material thickness d, and operating frequency f, and calculate the reflection loss RL value through the transmission line model;

[0011] Step (3): Based on the parallel computing strategy of tensor operation, use GPU hardware acceleration to generate a three-dimensional parameter space of RL values, and discretize the parameter space with engineering-level accuracy;

[0012] Step (4): Implement a dual-task screening strategy:

[0013] Task 1: Extract the dielectric constant (ε, ε) combination corresponding to the minimum RL under each (d, f) combination;

[0014] Task 2: Extract the boundary of the effective region of the dielectric constant corresponding to different RL thresholds;

[0015] Step (5): Based on the dual-task screening data set obtained in Step (4), construct a neural network-based dielectric constant prediction model. Input d and f, and output the predicted ε and ε. The model ensures the rationality of the output parameters through physical constraint optimization;

[0016] Step (6): Invert the predicted dielectric constant into the transmission line model to verify the wave-absorbing performance, generate the complete solution set distribution under the refined parameter space, and guide the customized design of magnetic wave-absorbing materials.

[0017] In the above solution, the magnetic wave-absorbing material uses carbonyl iron, and the above Step (1) includes:

[0018] Step (1.1): Prepare carbonyl iron magnetic materials with different rotation speeds through a ball milling process, and control the ball milling rotation speed in the range of 200 - 500 rpm;

[0019] Step (1.2): Measure the complex magnetic permeability parameter μ r = μ - jμ″ of the carbonyl iron material in the frequency band of 1 - 18 GHz;

[0020] Step (1.3): Use the measured complex magnetic permeability data sets obtained under the rotation speed conditions of 200 rpm and 500 rpm as the upper and lower boundary conditions of the magnetic permeability parameters.

[0021] In the above solution, the above Step (2) specifically includes:

[0022] (2.1) Discretize the material thickness d in the range of [0.5, 5.0] mm at intervals of 0.5 mm;

[0023] (2.2) Discretize the real part ε of the dielectric constant in the range of [1.0, 120.0] with a step size of 0.5, and discretize the imaginary part ε in the range of [0.00, 6.00] with a step size of 0.01;

[0024] (2.3) Calculate the reflection loss RL value through the transmission line model, and its expression is:

[0025]

[0026] RL = 20log|(Z in - Z0) / (Z in + Z0)|

[0027] where Z0 is the free space impedance, c is the speed of light, f is the operating frequency, and μ r is the preset dispersion complex permeability parameter.

[0028] In the above solution, the step (3) specifically includes:

[0029] Step (3.1) Construct the frequency tensor with the discretized frequency parameter ω = 2πf where N f = 160;

[0030] Step (3.2) Construct the thickness tensor with the material thickness parameter d where N d = 10;

[0031] Step (3.3) Construct the complex dielectric tensor with the dielectric constant parameter ε′ - jε″ where N ε = 240×600;

[0032] Step (3.4) Perform three-dimensional tensor operations through the Einstein summation method to generate a complex impedance tensor with dimensions of [N f × N d × N ε :

[0033]

[0034] where {k, l, m}, {l, m} are the dummy indices common in the Einstein summation formula and have no specific physical meaning for tensor calculation;

[0035] Step (3.5) Execute the broadcast mechanism for parallel calculation to complete the calculation of the reflection loss RL value of the parameter combination.

[0036] In the above solution, the step (4) includes the following parallel processing procedures:

[0037] Step (4.1) Task 1 Screening Process:

[0038] For each (d, f) combination, along the dielectric constant ε r dimension, extract the ε, ε″ combination corresponding to the minimum reflection loss RL, and record the corresponding RL value;

[0039] Step (4.2) Task 2 Screening Process:

[0040] (a) For the RL threshold set [-10dB, -8dB, -6dB], generate Boolean mask matrices respectively to mark the effective absorption regions;

[0041] (b) For each (d, f, threshold) combination, extract the ε′, ε″ coordinate boundaries of the effective ε r region;

[0042] Step (c) Determine the maximum value ε′ max and minimum value ε′ min of ε′, as well as the maximum value ε″ max and minimum value ε″ min of ε″.

[0043] In the above solution, the said step (5) specifically includes:

[0044] Step (5.1) Construct a fully connected neural network with an input layer, two hidden layers, and an output layer, where:

[0045] The input layer receives the normalized parameters of the material thickness d and the operating frequency f:

[0046]

[0047] The first hidden layer is configured with 64 neurons, and the second hidden layer is configured with 32 neurons, both using the ReLU activation function;

[0048] The output layer generates the normalized dielectric constant prediction value:

[0049]

[0050] where X is the input features d and f, and y corresponds to the output parameters ε′ and ε″;

[0051] Step (5.2) Apply physical constraints to the output layer through the softplus activation function;

[0052] Step (5.3) Use the Adam optimizer to perform 500 rounds of training with an initial learning rate of 0.001, the batch size n is 64, and the loss function is:

[0053]

[0054] In step (5.4), the validation set loss is monitored through a dynamic early stopping strategy, and the training is terminated when there is no improvement for p consecutive rounds, where p is set to 50.

[0055] The method for designing microwave absorbing materials guided by neural networks provided by the present invention significantly improves the optimization efficiency and design accuracy of magnetic microwave absorbing materials by integrating physical models and deep learning technologies, specifically embodying the following technical advantages:

[0056] 1. Decoupled optimization of magnetic and electrical parameters and improvement of computational efficiency

[0057] Through the decoupled design paradigm of "magnetic permeability locking - dielectric optimization", the dispersion characteristics of magnetic materials are fixed as boundary conditions, effectively reducing the complexity of multi-parameter coupled optimization.

[0058] Adopting a parallel computing strategy based on tensor operations and combining GPU hardware acceleration technology, rapid calculation of a three-dimensional parameter space (230.4 million combinations) is achieved (for example, it only takes 10 seconds in the embodiment), which is thousands of times more efficient than the traditional CPU nested loop, breaking through the bottleneck of calculating a large number of parameter combinations.

[0059] 2. Dual-task screening strategy for optimizing design objectives

[0060] In Task 1, by extracting the dielectric constant combinations corresponding to the minimum reflection loss (RL) under each (d, f) combination, the optimal microwave absorption performance parameters are accurately located;

[0061] In Task 2, by extracting the boundaries of the effective dielectric constant regions for different RL thresholds (-10 dB, -8 dB, -6 dB), a quantitative basis is provided for broadening the effective absorption bandwidth (EAB), realizing the multi-objective collaborative optimization of microwave absorption performance.

[0062] 3. High accuracy and physical rationality of the neural network model

[0063] The constructed fully connected neural network model dynamically constrains the physically reasonable range of the dielectric constant by introducing the softplus activation function, avoiding the generation of non-physical solutions (for example, eliminating negative ε values in the embodiment).

[0064] The R2 of the validation set of the model reaches 0.9889 (Task 1) and 0.9977 (Task 2) under the condition of 200 rpm, and the prediction results are highly consistent with the true values (as shown in Figure 6 、 7 、Figure 13), and the microwave absorption performance (RL < -20 dB) is verified by the transmission line model inversion, forming a "prediction - verification" closed loop to ensure the design reliability.

[0065] 4. Ability to guide the complete solution set of the refined parameter space

[0066] Generate the full-space solution set of dielectric constants at a refined thickness (0.1 mm) through neural network inference (such as Figure 10 , 14 ), visually reveal the coupling relationship between material thickness, frequency, and electromagnetic parameters, and provide high-resolution data support for customized design.

[0067] Combined with the prediction of the boundary of the effective absorption region, achieve rapid reverse design of material properties, significantly shorten the R & D cycle, and reduce the cost of experimental trial and error.

[0068] 5. Advantages of universality and scalability

[0069] The magnetic permeability boundary conditions (200 - 500 rpm) constructed with carbonyl iron as an example cover the characteristics of typical magnetic materials, and the model can be extended to other magnetic material systems (such as ferrites, metal alloys).

[0070] The neural network architecture supports dynamic adjustment (such as the number of hidden layers, activation functions), adapts to different frequency bands, material types, and performance index requirements, and has good technology migration ability.

[0071] In summary, through the deep integration of physical mechanisms and data-driven approaches, the present invention solves the core problems of low efficiency, poor accuracy, and large design blindness in the optimization of electromagnetic parameters of magnetic wave-absorbing materials, and provides an innovative solution for the intelligent design of high-performance wave-absorbing materials. Description of the Drawings

[0072] Figure 1 Magnetic permeability under different rotational speed conditions, where (a) μ′, (b) μ″;

[0073] Figure 2 Example of Task 1 dataset;

[0074] Figure 3 Example of Task 2 dataset;

[0075] Figure 4 Neural network model architecture diagram; Figure 5 MSE curve and R2 score curve of the Task 1 dataset corresponding to 200 rpm;

[0076] Figure 6 For the training set, validation set, test set, and all datasets of Task 1 corresponding to 200 rpm ε Scatter plot of the comparison between the predicted values and the true values of v and ε″;

[0077] Figure 7 , Figure 15 , Figure 16 Comparison curves of the predicted values and the true values of ε and ε″ of the model predicting Task 1 corresponding to 200 rpm under different thickness conditions;

[0078] Figure 8 For task 1 corresponding to 200 rpm predicted by the model without physical constraints, there is a negative solution for the ε″ value of 1.0 mm;

[0079] Figure 9 Curve comparing the RL results and screening results calculated from the predicted values of task 1 corresponding to 200 rpm predicted by the model under different thickness conditions;

[0080] Figure 10 Full-space ε′ and ε″ predicted solutions for task 1 corresponding to 200 rpm reconstructed by model inference;

[0081] Figure 11 RL results calculated from the full-space ε′ and ε″ predicted solutions for task 1 corresponding to 200 rpm reconstructed by model inference;

[0082] Figure 12 MSE curve and R 2 score curve for dataset -10 labels of task 2 corresponding to 200 rpm;

[0083] Figure 13 Curve comparing the predicted upper and lower boundary values and the true values of ε′ of task 2 (-10 labels) corresponding to 200 rpm predicted by the model under different thickness conditions;

[0084] Figure 14 Full-space predicted solutions of the upper and lower boundaries of ε′ for task 2 (-10 labels) corresponding to 200 rpm reconstructed by model inference. Detailed implementation mode

[0085] The following will give a detailed description of the embodiments of the present invention. Although the present invention will be described and explained in conjunction with some specific implementation modes, it should be noted that the present invention is not limited only to these implementation modes. On the contrary, modifications or equivalent replacements made to the present invention shall be covered within the scope of the claims of the present invention.

[0086] In addition, for better explaining the present invention, numerous specific details are given in the following detailed implementation modes. Those skilled in the art will understand that the present invention can also be implemented without these specific details.

[0087] The present invention proposes a decoupled design paradigm of "permeability locking - dielectric optimization" to break through the problem of synchronous optimization of traditional electromagnetic parameters. Here, as a possible implementation mode, taking ball-milled carbonyl iron as an example, a high-quality dataset for predicting the microwave absorption performance of magnetic materials is constructed. Specifically, considering that the permeability of existing magnetic microwave absorption materials only shows a similar change trend and small fluctuations in the range of 1 - 18 GHz, the measured permeability of carbonyl iron ball-milled at 200 rpm and 500 rpm is selected as the reference boundary condition ( Figure 1in a-b). These values effectively cover the vast majority of permeability parameters in magnetic wave-absorbing materials and have good universality.

[0088] Based on the above-mentioned relative complex permeability with dispersion characteristics as a reference, a high-throughput screening framework based on tensor operations was constructed to establish the dielectric constant feature space, effectively combining the underlying physical mechanism and the data dual-driven modeling requirements. Using the transmission line model, four key variables were systematically integrated: the permeability μ r = μ′ - jμ″, the dielectric constant ε r = ε′ - jε″, the material thickness d, and the operating frequency f. Where μ′ is the real part of the permeability corresponding to 200 rpm or 500 rpm, μ is the imaginary part of the permeability corresponding to 200 rpm or 500 rpm, ε′ is the real part of the dielectric constant corresponding to 200 rpm or 500 rpm, ε" is the imaginary part of the dielectric constant corresponding to 200 rpm or 500 rpm, and j is the imaginary unit.

[0089] The parameter space was discretized with engineering-level precision:

[0090] d ∈ [0.5, 5.0], Δd = 0.5 mm (1)

[0091] ε′ ∈ [1.0, 120.0], Δε′ = 0.5 (2)

[0092] ε″ ∈ [0.00, 6.00], Δε″ = 0.01 (3)

[0093] This strategy ensures that each (d, f) combination can generate 1.44×106 (ε′, ε″) combinations, achieving a good balance between computational feasibility and physical resolution. Subsequently, the wave-absorbing performance was quantified by calculating the reflection loss (RL) value, where Z0 is the free space impedance and c is the speed of light:

[0094]

[0095] The PyTorch tensor processing paradigm was adopted to achieve hardware-accelerated parallel computing, and the three-dimensional tensor operation was used to replace the traditional nested loop through the broadcast mechanism:

[0096] Frequency tensor:

[0097] Thickness tensor:

[0098] Dielectric constant tensor:

[0099] The hyperbolic tangent parameter was calculated by the Einstein summation method to form a dimension of [N f × Nd ×N ε complex tensors. Where N f = 160, N d = 10, N ε = 240 × 600, the total parameter space contains 230.4 million combinations.

[0100] Where {k, l, m}, {l, m} are dummy indices conventioned in the Einstein summation formula and have no specific physical meaning for tensor calculations;

[0101] Implement a dual - task screening strategy for the generated RL results. Task 1 aims to obtain the optimal RL value, that is, for each (d, f) combination, extract the (ε′, ε″) combination corresponding to the minimum RL along the ε r dimension and record the corresponding RL value. At the same time, in engineering applications, a wider effective absorption bandwidth (EAB) is also of concern. Therefore, Task 2 focuses on extracting the effective ε r regions where RL is lower than different thresholds (-10, -8, -6). For each (d, f, threshold) triple, use a boolean mask to extract the coordinates corresponding to the effective εr region, and determine the upper and lower boundaries of ε′ and ε″ through boundary retrieval. Figure 2 and Figure 3 shows the dataset form of the dual - task screening strategy. The entire calculation only takes 10.076 seconds on an NVIDIA RTX 4070Ti GPU, achieving thousands of times of acceleration compared to traditional CPU nested loops.

[0102] Different from traditional non - magnetic absorbing materials, the dielectric constant dataset obtained based on the dual - task screening strategy presents complex non - linear characteristics, which poses a major challenge to traditional function fitting methods. The inherent complexity of its contour trend makes it difficult to achieve mathematical parameterization, and the fitted function form lacks a concise mathematical form, making it difficult to guide material design. As an effective strategy for solving deep - learning non - linear regression problems, neural networks provide a breakthrough solution idea for the effective management of complex electromagnetic parameters through inherent characteristics such as autonomous multi - level feature extraction and data - driven pattern recognition. This application innovatively applies this artificial intelligence paradigm to construct a dielectric constant fitting model based on neural networks for the reverse intelligent design and development of magnetic absorbing materials.

[0103] Based on the dataset of Task 1, first conduct the prediction of the dielectric constant corresponding to the optimal RL. Construct a neural network with 2 hidden layers, which has 2 input neurons (d and f) and 2 output neurons (predicted ε′ and ε″). The network architecture adopts a fully - connected configuration, and the number of neurons in the 2 hidden layers is 64 and 32 respectively. Use the ReLU activation function to enhance the non - linear mapping ability of the model. The model architecture diagram is as Figure 4As shown in the figure, the dataset is divided into a training set, a validation set, and a test set according to a ratio of 7:2:1. Standard normalization is performed through z-score transformation to ensure data stability and accelerate model convergence:

[0104]

[0105] where X are the input features d and f, and y corresponds to the output parameters ε′ and ε″, and σ X represents the standard deviation of the input feature dataset, and μ X represents the mean of the input feature dataset, and σ y represents the standard deviation of the output parameter dataset, and μ y represents the mean of the output parameter dataset. In terms of model architecture design, a hyperparameter optimization scheme was obtained through parameter space exploration. The key parameters configured in the training protocol are as follows: the batch size is 64, the initial learning rate is 0.001, the number of training epochs is 500, and the Adam optimizer is used to achieve adaptive learning rate optimization in the complex parameter space. To quantify the model accuracy, the mean squared error (MSE) was defined as the loss function:

[0106]

[0107] where n is the batch size, and where represents the predicted value corresponding to the standardized true value ε′ of sample i, where i and where represents the predicted value corresponding to the standardized true value ε″ of sample i i At the same time, the training process was optimized through a dynamic early stopping strategy to prevent the model from overfitting. In addition, to improve the generalization ability of the model and address the physical rationality problem of data distribution extrapolation, the basic physical constraints were incorporated into the model architecture through the softplus activation transformation. This innovative method can achieve dynamic adjustment of the output boundary dominated by the statistical characteristics of the training dataset, while strictly restricting the predicted dielectric constant within the physically reasonable range. This mathematical non-negativity constraint ensures the robustness of the network architecture and realizes the internal integration of the physical principle and data dual-driven paradigm.

[0108] As Figure 5 shown, taking the task 1 dataset under the condition of 200 rpm as an example, the model demonstrates excellent performance: the MSE of the training set and the validation set continuously decreases with the increase in the number of training epochs, and the R2 of the validation set reaches 0.9889. The comparison scatter plot of the predicted value and the true value is as Figure 6 shown, showing stable prediction performance for both ε′ and ε″. When constructing the parameter space, the selected range of ε″ (0 - 6) is sufficient to reflect the characteristics of the actual material. At the same time, the model shows a slight deviation near the ε″ boundary. This weak edge effect can be ignored and does not affect the overall accuracy of the model.

[0109] like Figure 7 As shown in the figure, the model constrained by softplus shows excellent physical rationality, and the predicted value of the dielectric constant is in perfect agreement with the true value. Although the model without physical constraints also shows good prediction accuracy, the ε″ generated by it is negative, which violates the objective physical law ( Figure 8 ). More importantly, accurate prediction of dielectric constant is not the ultimate goal. The present invention then performs electromagnetic theoretical calculations. Converting the predicted ε and ε″ into RL results by integrating transmission line theory and judging whether the absorption performance has achieved the optimal effect is an important basis for verifying the effectiveness of the model. Figure 9 As shown in the figure, the RL value calculated based on the predicted dielectric constant is highly consistent with the screening result in the low frequency band, and achieves an ideal absorption effect in the high frequency band (RL<-20dB). In order to improve the practicality of material design, the thickness resolution is refined from 0.5mm to 0.1mm, and the neural network dielectric constant fitting model is used for inference in the full parameter space. The model has generated ε′, ε" and verified the complete solution set distribution of RL in the refined multidimensional space ( Figure 10 and Figure 11 ), visually analyzing the complex coupling relationship between the dielectric constant, geometric thickness, operating frequency and microwave absorption characteristics of magnetic carbonyl iron. This high-dimensional visual analysis establishes a quantitative framework for determining the optimal absorption region. It is worth noting that compared with traditional machine learning methods, this model combines physical information to form a closed-loop verification to ensure design feasibility and efficiency.

[0110] The fitting model architecture of the Task 2 data set is basically consistent with that of Task 1. Preliminary evaluation found that for the effective absorption areas corresponding to the three labels in Task 2, the preset range of ε″ (0 to 6) is almost fully satisfied, so only the upper and lower boundaries of ε′ need to be fitted. The data set of Task 2 is extracted according to the labels (-10, -8, -6), and the above-mentioned neural network dielectric constant fitting model is called 3 times to generate the ε′ solution space under different RL thresholds. On the basis of retaining the dual-output neuron structure of Task 1, the ε′ and ε″ result values previously used by the model to predict the ideal absorption conditions are replaced with the upper and lower boundary values of ε′ corresponding to the prediction of each effective absorption condition. It is worth noting that in order to avoid the abnormal phenomenon that the upper boundary value of ε′ is lower than the lower boundary value, the data sets corresponding to the two predicted values in Task 2 are jointly normalized, rather than independently normalized as in Task 1. The model exhibits excellent performance and strong generalization ability, with an R2 score of 0.9977 (0.9977) for the RL<-10dB target on the 200rpm validation set. Figure 12 ), the predicted ε′ boundary is highly consistent with the screening results ( Figure 13 ). Similarly, by reconstructing and refining the parameter space through reasoning, the full space results of the upper and lower boundaries of ε′ are obtained as follows Figure 14As shown, this lays a theoretical foundation for the prediction of each target effective absorption region and the customized design of microwave absorption performance. The model also performs well on the 500 rpm condition dataset, which will not be repeated here.

[0111] In summary, the collaborative application of the dual-task screening strategy developed in this patent and the neural network-guided magnetic microwave absorption material design model successfully maps the dielectric constant solution set space of ideal microwave absorption performance and effective microwave absorption performance, providing a more convenient and intelligent new strategy for the design of microwave absorption materials.

Claims

1. A design method of microwave absorbing materials guided by neural network, characterized in that, It includes the following steps: Step (1): Based on the magnetic permeability dispersion characteristics of the magnetic wave-absorbing material, select the measured magnetic permeability of the magnetic material with dispersion characteristics as the reference boundary condition; Step (2) constructs a high-throughput screening parameter space including the magnetic permeability μ r , the permittivity ε r , the material thickness d, and the operating frequency f, and calculates the reflection loss RL value through the transmission line model; Step (3): Based on the parallel computing strategy of tensor operation, use GPU hardware acceleration to generate the three-dimensional parameter space of RL values, and discretize the parameter space with engineering-level accuracy; Step (4): Implement a dual-task screening strategy: Task 1: Extract the permittivity (ε′, ε″) combinations corresponding to the minimum RL under each (d, f) combination; Task 2: Extract the boundaries of the effective permittivity regions corresponding to different RL thresholds; Step (5): Based on the dual-task screening data set obtained in Step (4), construct a neural network-based permittivity prediction model. Input d and f, and output the predicted ε′ and ε″. The model ensures the rationality of the output parameters through physical constraint optimization; Step (6): Invert the predicted permittivity into the transmission line model to verify the wave-absorbing performance, generate the complete solution set distribution in the refined parameter space, and guide the customized design of the magnetic wave-absorbing material.

2. The method for designing a microwave absorbing material guided by a neural network according to claim 1, wherein The magnetic wave-absorbing material uses carbonyl iron, and Step (1) includes: Step (1.1): Prepare carbonyl iron magnetic materials with different rotation speeds through the ball milling process, and control the ball milling speed in the range of 200 - 500 rpm; Step (1.2) measures the complex permeability parameter μ of the carbonyl iron material within the frequency band of 1 - 18 GHz r = μ′ - jμ″; Step (1.3): Use the measured complex magnetic permeability data sets obtained under the rotation speeds of 200 rpm and 500 rpm as the upper and lower boundary conditions of the magnetic permeability parameters.

3. The method for designing an electromagnetic wave absorbing material guided by a neural network according to claim 1, characterized in that Step (2) specifically includes: (2.1) Discretize the material thickness d in the range of [0.5, 5.0] mm at intervals of 0.5 mm; (2.2) Discretize the real part ε′ of the permittivity at intervals of 0.5 in the range of [1.0, 120.0], and discretize the imaginary part ε″ at intervals of 0.01 in the range of [0.00, 6.00]; (2.3) Calculate the reflection loss RL value through the transmission line model, and its expression is: RL = 20 log|(Z in - Z0) / (Z in + Z0)| where Z0 is the free space impedance, c is the speed of light, f is the operating frequency, and μ r is a preset dispersion complex permeability parameter.

4. The method for designing a microwave absorbing material guided by a neural network according to claim 1, characterized in that Step (3) specifically includes: Step (3.1) constructs a frequency tensor using the discretized frequency parameter ω = 2πf where N f = 160; Step (3.2) constructs the material thickness parameter d as a thickness tensor where N d = 10; Step (33) constructs the dielectric constant parameter ε′ - jε″ as a complex dielectric tensor where N ε = 240 × 600; Step (3.4) performs three-dimensional tensor operations through Einstein summation to generate a complex impedance tensor with dimensions of [N f ×N d ×N ε : Among them, {k, l, m}, {l, m} are the dummy indices agreed in the Einstein summation formula and have no specific physical meaning for tensor calculation; Step (3.5): Execute the broadcast mechanism for parallel calculation to complete the calculation of the reflection loss RL value of the parameter combination.

5. The method for designing a microwave absorbing material guided by a neural network according to claim 1, wherein Step (4) includes the following parallel processing processes: Step (4.1): Task 1 screening process: For each (d, f) combination, along the dielectric constant ε r Extract the combination of ε′ and ε″ corresponding to the minimum reflection loss RL in the r dimension, and record the corresponding RL value; Step (4.2): Task 2 screening process: (a) For the RL threshold set [-10dB, -8dB, -6dB], generate boolean mask matrices to mark the effective absorption regions respectively; (b) For each (d, f, threshold) combination, extract the effective ε r The ε′ and ε″ coordinate boundaries of the region; Step (c) determines the maximum value ε′ of ε′ and the minimum value ε′ of ε′, as well as the maximum value ε″ of ε″ and the minimum value ε″ of ε″ through a boundary retrieval algorithm. max and the minimum value ε′ min , as well as the maximum value ε″ of ε″ max and the minimum value ε″ min .

6. The method for designing a microwave absorbing material guided by a neural network according to claim 1, characterized in that Step (5) specifically includes: Step (5.1): Construct a fully connected neural network with an input layer, two hidden layers, and an output layer, where: The input layer receives the normalized parameters of the material thickness d and the operating frequency f: The first hidden layer is configured with 64 neurons, and the second hidden layer is configured with 32 neurons, both using the ReLU activation function; The output layer generates the normalized permittivity prediction values: Among them, X is the input features d and f, and y corresponds to the output parameters ε′ and ε″; Step (5.2): Apply physical constraints to the output layer through the softplus activation function; Step (5.3) uses the Adam optimizer to perform 500 rounds of training with an initial learning rate of 0.001, the batch size n is 64, and the loss function is: Step (5.4) monitors the validation set loss through a dynamic early stopping strategy and terminates the training when there is no improvement for p consecutive rounds.

Images