Gradient Absorbing Cellular Design Method Based on Neural Networks and Random Explosion Search

Abstract

This invention belongs to the field of electromagnetic wave absorbing materials technology, specifically relating to a gradient absorbing cell design method based on neural networks and random burst search. First, a gradient cell model is established in full-wave simulation software, and sufficient training data is obtained through simulation. Then, a high-performance neural network model is established and trained to achieve rapid mapping of gradient cell structure parameters to reflection loss curves. Finally, leveraging the rapid mapping capability of the trained neural network, large-scale random search and small-scale burst search are combined to solve the multi-objective design problem of maximizing absorption bandwidth and maximizing absorption performance within a specific frequency band in the gradient cell absorbing material design process. This invention has the advantages of saving computational resources, high design efficiency, and good multi-objective optimization effect. It optimizes bandwidth while considering absorption performance in a specific frequency band, enabling more accurate and user-friendly optimized designs.

Description

Technical Field

[0001] This invention belongs to the field of electromagnetic wave absorbing materials technology, specifically relating to a gradient absorbing cell design method based on neural networks and random explosion search. Background Technology

[0002] In today's era of widespread 5G, while enjoying the convenience brought by modern communication tools and electronic products, people are also exposed to an environment with increasing electromagnetic radiation. Everyday electronic devices and household appliances have become "invisible killers" of indoor electromagnetic radiation. The large amount of electromagnetic radiation has caused electromagnetic pollution in the environment. This not only greatly affects people's daily lives but also poses a significant threat to human health. Therefore, how to effectively reduce electromagnetic radiation pollution has become a hot research topic in today's society.

[0003] Gradient honeycomb absorbing materials are highly efficient materials for electromagnetic wave absorption. They typically consist of a multi-layered honeycomb structure, with each layer exhibiting different absorption properties, displaying a gradient distribution. This gradient design helps optimize the material's absorption performance, especially within specific frequency bands. Compared to single-layer honeycomb materials, gradient honeycomb structures can effectively reduce electromagnetic wave reflection and signal interference; the gradient design also extends the operating frequency band of the absorbing material, enabling it to perform well over a wider frequency range.

[0004] The design of gradient cellular absorbing materials is a complex multi-objective optimization problem that requires balancing multiple objectives to obtain optimal performance. The traditional design process for gradient cellular absorbing materials involves complex electromagnetic principles and models, typically requiring the use of computer-aided design tools and simulation software, combined with commonly used multi-objective optimization algorithms such as genetic algorithms, particle swarm optimization, and ant colony optimization, which consumes a great deal of computational resources and time. Summary of the Invention

[0005] To address the aforementioned problems and shortcomings, and to resolve the issues of high computational resource requirements and extreme time dependence on design time in existing gradient cellular absorbing material design processes, this invention provides a gradient absorbing cellular design method based on neural networks and random burst search, which can achieve or even surpass the design results of traditional methods in a very short time. This invention constructs a neural network to predict the absorption performance of gradient cellular absorbing materials using their structural parameters. Furthermore, leveraging the rapid mapping capability of the neural network, it combines large-scale random search with small-scale burst search to achieve multi-objective design of gradient cellular absorbing materials based on neural networks. This solves the multi-objective design problem of maximizing absorption bandwidth and maximizing absorption performance within a specific frequency band in the design of gradient cellular absorbing materials under constraints on total height.

[0006] The technical solution adopted in this invention is as follows:

[0007] A gradient-absorbing cell design method based on neural networks and random explosion search includes the following steps:

[0008] Step 1: Use full-wave simulation software to model and simulate a gradient honeycomb structure with a certain number of layers, and establish a dataset showing the relationship between the structural parameters of the gradient honeycomb absorbing material and the corresponding reflection loss curve.

[0009] Step 2: Establish a neural network model. By learning and simulating the inherent patterns of the data in the training dataset, the structural parameters of the gradient honeycomb absorbing material can be quickly mapped to the reflection loss curve.

[0010] Step 3: Combine the trained neural network with large-scale random search and small-scale explosive search for multi-objective optimization design of gradient cellular absorbing materials.

[0011] Step 3-1. Set optimization conditions: including the frequency range for wideband optimization (Freq). min ,Freq max Gradient cell total height limit H lim Target reflection loss (RL) tar The focus is on optimizing the frequency range (freq). min ,freq max ) and sample screening time limit t lim .

[0012] Step 3-2. Randomly generate parameter combinations within the variable parameter range, input them into the neural network to obtain the corresponding reflection loss curves, and select l0 (l0≥1000) reflection loss curves that meet the optimization conditions. Calculate their bandwidth BW and the average reflection loss RL within the key optimization frequency band. Avg ;

[0013] RL i This represents the reflection loss corresponding to the i-th frequency point within the broadband optimized frequency band. Where Z represents the set of integers. The number of frequency points within the wideband optimization band where the reflection loss is better than the target reflection loss is... Bandwidth corresponding to the reflection loss curve

[0014] RL j This indicates that the focus is on optimizing the reflection loss at the j-th frequency point within the frequency band. The corresponding reflection loss curve shows the average reflection loss within the key optimization frequency band.

[0015] Step 3-3. Calculate the bandwidth BW and average reflection loss RL of the 10 reflection loss curves that satisfy the optimization conditions obtained in Step 3-2. Avg The absolute values ​​are weighted and summed according to the following formula, which serves as the performance index of the reflection loss curve:

[0016] ∑ std = (1-a)·BW+a·|RL Avg |

[0017] Where a (0≤a≤0.1) is a constant and can be adjusted according to the optimization conditions, Σ std The larger the value, the better the performance of the reflection loss curve.

[0018] Calculate Σ for each reflection loss curve std Sort the curves in descending order and retain the top n reflection loss curves with the best performance. For each reflection loss curve, perform an explosive search on the k parameters (expanding the range to 3 by adding and subtracting within x times the minimum unit value). k The new parameters are set and input into the neural network to predict the new reflection loss curve (x≤5), and the corresponding Σ is calculated. std The best-performing reflection loss curve is retained. Following this process, n reflection loss curves correspond to n optimized new reflection loss curves, which are used for the next round of optimization.

[0019] Steps 3-4. Σ of the n new reflection loss curves after the previous round of optimization. std If the maximum value fluctuates downwards by no more than 5%, update the optimization conditions. Repeat step 3-2, and use the neural network to continue filtering l (l≤0.1×l0) reflection loss curves that meet the updated optimization conditions. Combine these with the reflection loss curves optimized in the previous round and perform another explosive search. Repeat step 3-3 to select the n best-performing reflection loss curves from the expanded reflection loss curves.

[0020] Step 3-5. Repeat step 3-4. If the time it takes for the neural network to select a reflection loss curve that meets the optimization conditions exceeds the set time t... lim If the current performance fails, the selection process stops. Instead, the best n existing reflection loss curves are optimized through multiple rounds of explosive search until the total cell height limit prevents the generation of any further better reflection loss curves. The structural parameters corresponding to the best-performing reflection loss curve obtained at this point are the final design parameters.

[0021] Furthermore, step 1 specifically includes:

[0022] Step 1-1. Determine the structural parameters of the gradient honeycomb absorbing material. In full-wave simulation software (such as CSTMicrowaveStudio), establish the corresponding periodic structure model of the gradient honeycomb absorbing material, set the simulation frequency range, and import the electromagnetic parameters of the absorbing paste into the model.

[0023] Step 1-2. Set the k types of structural parameters that actually need to be optimized during the design process as variable parameters, and determine the range and minimum unit value of each variable parameter. Set the remaining structural parameters as fixed parameters.

[0024] Steps 1-3. Randomly generate N (N≥5000) combinations of variable parameters and input them into the full-wave simulation software for simulation calculation to obtain the corresponding reflection loss curves. From each reflection loss curve, take the reflection loss at the corresponding frequency point at an interval of p (p≤0.25GHz) to form a set, and save it together with the k variable parameters corresponding to that reflection loss curve as a text file as a set of training data for the neural network. The entire training dataset contains a total of N sets of training data.

[0025] Furthermore, step 2 specifically includes:

[0026] Step 2-1. Divide the N sets of data obtained in Step 1-3 into training, validation, and test sets according to a ratio (usually between 6:2:2 and 8:1:1). Build a neural network using an open-source deep learning framework, using k variable parameters as input parameters and the corresponding reflection loss set as the target parameters. Configure the optimizer and loss function type during neural network training, set the initial learning rate and maximum training epochs, and lock the random seed to ensure reproducibility of the training process.

[0027] Step 2-2. To monitor the performance of the neural network in real time during training, a custom accuracy rate is introduced: if the error between the reflection loss output by the neural network and the simulation result at the corresponding frequency point does not exceed 5%, it is considered an accurate prediction. The proportion of accurately predicted frequencies to the total number of frequencies on all reflection loss curves during a training round is considered the accuracy rate for that round of training.

[0028] Begin neural network training by extracting input and target parameters from the training set and training the neural network. After each training epoch, use the validation set to evaluate the model's performance to aid in model tuning. Save the loss function and accuracy of the training and validation sets after each training epoch.

[0029] Steps 2-3. Based on the trends in the loss function and prediction accuracy during neural network training (ensuring the neural network model does not suffer from underfitting due to insufficient training, or overfitting due to excessive training leading to low prediction accuracy), determine the optimal number of training epochs, m, to stop the neural network training. Reproduce the neural network training process, setting the training to terminate after m epochs, to obtain the trained neural network.

[0030] Steps 2-4. Perform performance testing on the trained neural network using the test set, defining an accuracy metric: if the percentage of accurately predicted frequencies on a reflection loss curve is greater than or equal to 95%, then that reflection loss curve is considered accurately predicted. In one round of testing, if the percentage of accurately predicted reflection loss curves is greater than 95%, the neural network performance is considered satisfactory and can be used for subsequent optimization.

[0031] This invention leverages the rapid development of artificial intelligence algorithms, fully utilizing their ability to simulate highly complex nonlinear mappings, thereby simplifying or circumventing the challenges of originally complex and difficult-to-calculate physical models. First, this invention uses artificial intelligence to construct a neural network to predict the absorption performance of gradient honeycomb absorbing materials based on their structural parameters. Furthermore, by utilizing the rapid predictive capabilities of the neural network, it combines large-scale random search with small-scale explosive search to achieve multi-objective optimization design of gradient honeycomb absorbing materials based on neural networks.

[0032] In summary, this invention, by introducing a fast and highly accurate neural network, eliminates the need for extensive simulation calculations in the design process, thus improving design efficiency and enabling large-scale random searches and small-scale explosive searches. It can optimize bandwidth while also considering the absorption performance of specific frequency bands, achieving more precise and user-friendly multi-objective optimization designs. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the process of the present invention.

[0034] Figure 2 This is a schematic diagram of the structure of the three-layer gradient honeycomb absorbing material in Example 3.

[0035] Figure 3 This is a schematic diagram of the neural network structure for an example.

[0036] Figure 4 The example shows the accuracy and loss function curves of the training and validation sets during neural network training.

[0037] Figure 5 This example demonstrates the effect of comparing the neural network-predicted reflection loss curve with the simulated reflection loss curve.

[0038] Figure 6This is a comparison chart of the reflection loss curves of the example and the structure optimized only for broadband. Detailed Implementation

[0039] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0040] A gradient absorbing cell design method based on neural networks and random explosion search (e.g.) Figure 1 (As shown), including the following steps:

[0041] Step 1: Use the full-wave simulation software CST to model and simulate the gradient honeycomb structure with a certain number of layers, and establish a dataset showing the relationship between the structural parameters of the gradient honeycomb absorbing material and the corresponding reflection loss curve.

[0042] Step 1-1. Establish a periodic structure model of a three-layer gradient honeycomb absorbing material in the full-wave simulation software CST Microwave Studio (e.g., Figure 2 As shown in the figure, the simulation frequency range was set to 0.5-18GHz, and the electromagnetic parameters of the absorbing slurry were imported into the model.

[0043] Step 1-2. Set a total of 6 variable parameters (h1, h2, h3, t1, t2, t3) for the honeycomb height and slurry thickness of each layer in the three-layer gradient honeycomb absorbing material. The honeycomb height h of each layer from the outside to the inside... i =0.05×h0, {x|80≤h0≤200,x∈N}, i=1,2,3, with a corresponding minimum unit value of 0.05mm. The thickness T of the slurry for each impregnation. i =0.0025×T0,{x|4≤T0≤40,x∈N},i=1,2,3, the thickness of the first layer of slurry from the outside to the inside is t1=T1, the thickness of the second layer of slurry is t2=T1+T2, and the thickness of the third layer of slurry is t3=T1+T2+T3, with a corresponding minimum unit value of 0.0025mm. Fixed parameters include honeycomb pore size d=2.75mm and honeycomb wall thickness w=0.05mm.

[0044] Steps 1-3. Randomly generate N = 30,000 combinations of variable parameters and input them into CST for simulation to obtain the corresponding reflection loss curves. From each reflection loss curve, select a set of 351 frequency points corresponding to the reflection loss at intervals of p = 0.05 GHz. Save this set along with the 6 variable parameters corresponding to that reflection loss curve as a text file as a set of training data for the neural network. The entire training dataset contains 30,000 sets of training data.

[0045] Step 2: Establish a convolutional neural network model. By learning and simulating the inherent patterns of the data in the training dataset, the structural parameters of the gradient honeycomb absorbing material can be quickly mapped to the reflection loss curve.

[0046] Step 2-1. Divide all 30,000 data sets obtained in Steps 1-3 into training, validation, and test sets in an 8:1:1 ratio. Use six variable parameters as input parameters for the neural network, and the corresponding reflection loss set as the target parameters. Build a convolutional neural network using the PyTorch open-source deep learning framework. The network structure is as follows: Figure 3 As shown. The training process uses the Adam optimizer, the mean squared loss function MSELoss, the initial learning rate is set to 0.001, the maximum number of training epochs is set to 1500, and a random seed is locked to make the training process reproducible.

[0047] Step 2-2. To monitor the performance of the neural network in real time during training, a custom accuracy rate is introduced: if the error between the reflection loss output by the neural network and the simulation result at the corresponding frequency point does not exceed 5%, it is considered an accurate prediction. The proportion of accurately predicted frequencies to the total number of frequencies on all reflection loss curves during a training round is considered the accuracy rate for that round of training.

[0048] Begin neural network training by extracting input and target parameters from the training set and training the neural network. After each training epoch, use the validation set to evaluate the model's performance to assist in model tuning. Save the loss function values ​​and accuracy of the training and validation sets after each training epoch.

[0049] Steps 2-3. The trends of loss function and accuracy during neural network training are as follows: Figure 4 As shown, the optimal number of epochs, m, to stop the neural network training is determined to be 982 epochs. The training process of the neural network is reproduced, with the training set to terminate after 982 epochs, resulting in a fully trained neural network.

[0050] Steps 2-4. Perform performance testing on the trained neural network using the test set, defining an accuracy metric: if the percentage of accurately predicted frequencies on a reflection loss curve is greater than or equal to 95%, then that reflection loss curve is considered accurately predicted. In one round of testing, if the percentage of accurately predicted reflection loss curves is greater than 95%, the neural network performance is considered satisfactory and can be used for subsequent optimization.

[0051] In this embodiment, the neural network prediction effect is as follows: Figure 5 As shown, the performance of the trained neural network was tested using 500 random sets of data from the test set. The average accuracy across 5 tests was over 97%, with a minimum accuracy of 95.73%, which can be used for subsequent optimization.

[0052] Step 3: Combine the trained neural network with large-scale random search and small-scale explosive search for multi-objective optimization design of gradient cellular absorbing materials.

[0053] Step 3-1. Set the frequency range for broadband optimization to 1-18GHz, and the total height limit for gradient cells to H. lim =20mm, target reflection loss RL tar = -10dB, with a focus on optimizing the frequency range of 6-12GHz, and a sample screening time limit of t. lim =180s.

[0054] Step 3-2. Randomly generate parameter combinations within the variable parameter range, input them into the neural network to obtain the corresponding reflection loss curves, and select l0 = 2500 reflection loss curves that meet the optimization conditions. Calculate their bandwidth BW and the average reflection loss RL within the key optimization frequency band. Avg .

[0055] Step 3-3. Calculate the bandwidth BW and average reflection loss RL of the 2500 reflection loss curves that meet the optimization conditions obtained in Step 3-2. Avg The absolute values ​​are weighted and summed according to the following formula, which serves as the performance index of the reflection loss curve:

[0056] ∑ std = (1-a)·BW+a·|RL Avg |

[0057] Taking a = 0.05, calculate the Σ for each reflection loss curve. std Arranged in descending order, the first 10 reflection loss curves are selected, along with their corresponding variable parameter combinations. For each reflection loss curve, the k=6 parameters are added to or subtracted from a minimum unit value (honeycomb height ±0.05mm, slurry thickness ±0.0025mm), and then combined to expand to 3. 6 New parameters are set and input into the neural network to predict and obtain a new reflection loss curve. The corresponding Σ is then calculated. std The best-performing reflection loss curve is then selected and retained. Following this process, 10 optimized new reflection loss curves are obtained for use in the next round of optimization.

[0058] Steps 3-4. Σ of the 10 reflection loss curves after the previous round of optimization. std Based on the maximum value, the optimization conditions are updated by floating down by 1%. Step 3-2 is repeated to use the neural network to select l=100 reflection loss curves that meet the updated optimization conditions, and these curves are then subjected to another explosive search together with the reflection loss curves optimized in the previous round. Step 3-3 is repeated to select the 10 best-performing reflection loss curves from the expanded reflection loss curves.

[0059] Steps 3-5. Repeat steps 3-4. If the time taken for the neural network to select a reflection loss curve that meets the optimization conditions exceeds 180 seconds, stop the selection process. Only optimize the existing top 10 reflection loss curves through multiple rounds of explosive search until the total cell height limit prevents the generation of any better reflection loss curves. At this point, the optimization process ends. The structural parameters corresponding to the best-performing reflection loss curve obtained at this time are the final design parameters.

[0060] Figure 6 In this embodiment, within a wide bandwidth optimization range of 1-18 GHz, the target reflection loss is set to -10 dB, and the total height of the gradient cell is ≤20 mm. The design results obtained by focusing on optimizing the 6-12 GHz frequency band are compared with the reflection loss curves corresponding to the four sets of structures obtained by optimizing only the wide bandwidth. The entire optimization process takes only about 15 minutes.

[0061] The structural parameters and wave absorption performance corresponding to curves ①②③ are as follows:

[0062] ① d = 2.75, w = 0.05, h1 = 6.78, h2 = 6.03, h3 = 7.18, t1 = 0.040, t2 = 0.105, t3 = 0.173, optimized bandwidth BW = 16.25dB, with a focus on optimizing the average reflection loss RL in the frequency band. Avg = -20.24dB, corresponding to Σ std =16.4495dB;

[0063] ②d=2.75,w=0.05,h1=6.73,h2=6.43,h3=6.00,t1=0.040,t2=0.113,t3=0.201, optimized bandwidth BW=16.25dB, focusing on optimizing the average reflection loss RL of the frequency band. Avg = -20.44dB, corresponding to Σ std =16.4595dB;

[0064] ③ d = 2.75, w = 0.05, h1 = 8.49, h2 = 5.43, h3 = 5.48, t1 = 0.056, t2 = 0.139, t3 = 0.159, optimized bandwidth BW = 16.25dB, with a focus on optimizing the average reflection loss RL in the frequency band. Avg = -19.68dB.

[0065] Figure 6 Among the four reflection loss curves optimized only for bandwidth, although reflection loss curve ③ has better overall absorption performance, its absorption performance in the 6-12GHz frequency band is weaker than that of reflection loss curves ① and ② obtained by multi-objective optimization.

[0066] As can be seen from the design results of the above embodiments, this invention first establishes a gradient honeycomb model in full-wave simulation software and obtains sufficient training data through simulation; then, by establishing and training a high-performance neural network model, it achieves rapid mapping from gradient honeycomb structure parameters to reflection loss curves; finally, by utilizing the trained neural network and its rapid mapping capability, it combines large-scale random search and small-scale explosive search to solve the multi-objective design problem of maximizing absorption bandwidth and maximizing absorption performance within a specific frequency band in the design of gradient honeycomb absorbing materials. This invention has the advantages of saving computational resources, high design efficiency, and good multi-objective optimization effect. It optimizes bandwidth while taking into account absorption performance in a specific frequency band, achieving a more accurate and user-friendly optimized design.

Claims

1. A gradient-absorbing cell design method based on neural networks and random explosion search, characterized in that, Includes the following steps: Step 1: Use full-wave simulation software to model and simulate a gradient honeycomb structure with a certain number of layers, and establish a dataset showing the relationship between the structural parameters of the gradient honeycomb absorbing material and the corresponding reflection loss curves. Step 2: Establish a neural network model. By learning and simulating the inherent patterns of the data in the training dataset, the structural parameters of the gradient honeycomb absorbing material can be quickly mapped to the reflection loss curve. Step 3: Combine the trained neural network with large-scale random search and small-scale explosive search for multi-objective optimization design of gradient cellular absorbing materials. Step 3-1. Set optimization conditions: including the frequency range for wideband optimization (Freq). min ,Freq max Gradient cell total height limit H lim Target reflection loss RL tar The focus is on optimizing the frequency range (freq). min ,freq max ) and sample screening time limit t lim ; Step 3-2. Randomly generate parameter combinations within the variable parameter range, input them into the neural network to obtain the corresponding reflection loss curves, and select l0 reflection loss curves that meet the optimization conditions, where l0≥1000. Calculate their bandwidth BW and the average reflection loss RL within the key optimization frequency band. Avg ; RL i This represents the reflection loss corresponding to the i-th frequency point within the broadband optimized frequency band. Where Z represents the set of integers; the number of frequency points in the wideband optimization band where the reflection loss is better than the target reflection loss is: Bandwidth corresponding to the reflection loss curve RL j This indicates that the focus is on optimizing the reflection loss at the j-th frequency point within the frequency band. The corresponding reflection loss curve shows the average reflection loss within the key optimization frequency band. Step 3-3. Calculate the bandwidth BW and average reflection loss RL of the 10 reflection loss curves that satisfy the optimization conditions obtained in Step 3-2. Avg The absolute values ​​are weighted and summed according to the following formula, which serves as the performance index of the reflection loss curve: ∑ std =(1-a)·BW+a·|RL Avg | Where a is a constant, 0 ≤ a ≤ 0.1, Σ std The larger the value, the better the performance of the reflection loss curve; Calculate Σ for each reflection loss curve std Sort the curves in descending order and retain the top n reflection loss curves with the best performance. Perform an explosive search on the k parameters of each reflection loss curve to calculate the corresponding Σ. std The best-performing reflection loss curve is retained; n reflection loss curves correspond to n optimized new reflection loss curves, which are used for the next round of optimization. The explosive search refers to expanding to 3 by performing addition and subtraction operations within a minimum unit value of x times. k The new parameters are set and input into the neural network to predict the new reflection loss curve, where x≤5; Steps 3-4. Σ of the n new reflection loss curves after the previous round of optimization. std If the maximum value fluctuates downward by less than 5%, update the optimization conditions; repeat step 3-2, and use the neural network to continue to select l reflection loss curves that meet the updated optimization conditions, l≤0.1×l0, and perform an explosive search again together with the reflection loss curves optimized in the previous round; repeat step 3-3, and select the n reflection loss curves with the best performance from the expanded reflection loss curves. Step 3-5. Repeat step 3-4. If the time it takes for the neural network to select a reflection loss curve that meets the optimization conditions exceeds the set time t... lim If the selection fails, the remaining selection process stops. Instead, the existing n best-performing reflection loss curves are optimized through multiple rounds of explosive search until the total cell height limit prevents the generation of any better-performing reflection loss curves. At this point, the optimization process ends. The structural parameters corresponding to the best-performing reflection loss curve obtained at this time are the final design parameters.

2. The gradient absorbing cell design method based on neural networks and random explosion search as described in claim 1, characterized in that, The specific details of step 1 are as follows: Step 1-1. Determine the structural parameters of the gradient honeycomb absorbing material, establish the corresponding periodic structure model of the gradient honeycomb absorbing material in the full-wave simulation software, set the simulation frequency range, and import the electromagnetic parameters of the absorbing paste into the model; Step 1-2. Set the k types of structural parameters that actually need to be optimized during the design process as variable parameters, and determine the range and minimum unit value of each variable parameter. Set the remaining structural parameters as fixed parameters. Steps 1-3. Randomly generate N combinations of variable parameters, N≥5000, input them into the full-wave simulation software for simulation calculation, and obtain the corresponding reflection loss curves; From each reflection loss curve, take the reflection loss at the corresponding frequency point at an interval of p to form a set, and save it together with the k variable parameters corresponding to that reflection loss curve as a text file as a set of training data for the neural network, where p≤0.25GHz; the entire training dataset contains N sets of training data.

3. The gradient absorbing cell design method based on neural networks and random explosion search as described in claim 1, characterized in that, The specific details of step 2 are as follows: Step 2-1. Divide the N sets of data obtained in Step 1-3 into training set, validation set and test set in a ratio of 6:2:2-8:1:1; build a neural network using an open-source deep learning framework, using k variable parameters as input parameters of the neural network, and the corresponding reflection loss set as target parameters of the neural network; set the optimizer and loss function type in the neural network training process, set the initial learning rate and maximum number of training epochs, and lock the random seed to make the training process reproducible; Step 2-2. In order to monitor the performance of the neural network in real time during training, a custom accuracy rate is introduced: if the error between the reflection loss output by the neural network and the simulation result of the corresponding frequency point does not exceed 5%, it is considered to be accurate; the proportion of accurately predicted frequency points to the total number of frequency points on all reflection loss curves in a training round is considered to be the accuracy rate of this training round. Start neural network training by extracting input and target parameters from the training set and training the neural network. After each training round, use the validation set to evaluate the model's performance and assist in model tuning. Save the loss function and accuracy of the training set and validation set after each round of training; Steps 2-3. Based on the changing trends of the loss function and prediction accuracy during the neural network training process, and on the premise that the neural network model will not underfit due to insufficient training or overfit due to excessive training, resulting in low prediction accuracy, determine the optimal number of rounds m to stop the neural network training; reproduce the neural network training process, set the neural network to terminate after m rounds of training, and obtain the trained neural network. Steps 2-4. Use the test set to test the performance of the trained neural network and define the accuracy index: if the percentage of accurately predicted frequency points on a reflection loss curve is greater than or equal to 95%, then the reflection loss curve is considered to be accurately predicted. If more than 95% of the predicted reflection loss curves are accurate in a single test, the neural network is considered to have met the performance standards and can be used for subsequent optimization.

4. The gradient absorbing cell design method based on neural networks and random explosion search as described in claim 1, characterized in that: The full-wave simulation software used is CST Microwave Studio.

5. The gradient absorbing cell design method based on neural networks and random explosion search as described in claim 1, characterized in that: The neural network model is a convolutional neural network built using the PyTorch open-source deep learning framework. The Adam optimizer is used during the training process, and the mean squared error (MSELoss) is used as the loss function.

Images