A multi-parameter broadband transparent wave-absorbing body design method based on a deep learning model

Abstract

This invention discloses a design method for multi-parameter broadband transparent absorbers based on a deep learning model. The method first constructs an absorbing unit composed of parameters such as a top-level metal encoding matrix, medium thickness, and unit period. Representative samples are selected using the Kennard-Stone algorithm for simulation to build a dataset. Subsequently, a Vision Transformer prediction model is established, achieving multi-modal feature fusion through graph embedding and scalar fully connected branches to predict the absorption rate curve. To address the convergence challenge of multi-parameter absorbers, an inheritance-based progressive training strategy is adopted, gradually releasing variables and incorporating historical samples for training. Finally, the trained model is used as a surrogate model in conjunction with a genetic algorithm for global optimization. This invention improves the design efficiency of complex multi-parameter transparent absorbers, achieves high-performance absorption across a wide bandwidth, and solves the problems of long design cycles and susceptibility to local optima in traditional designs.

Description

Technical Field

[0001] This invention relates to the fields of deep learning and transparent absorbers, specifically to a broadband multi-parameter input coded transparent absorber absorptivity prediction model. Background Technology

[0002] Transparent absorbers are a novel type of material that is both optically transparent and microwave-absorbing. At the beginning of this century, researchers successfully constructed Salisbury screens and Jaumann absorbers using stacked transparent conductive thin films, achieving both visible light transparency and microwave absorption. However, uniform thin film structures suffer from narrow absorption bandwidth, large thickness, and low visible light transmittance. By constructing microstructures on the thin film surface, the fundamental physical characteristics of electromagnetic waves, such as frequency, amplitude, phase, and polarization, can be controlled. Recent research has shown that transparent absorbers designed using metasurfaces have improved key performance aspects. However, existing metasurface designs often employ simple structures such as squares, circles, and rings, and perform electromagnetic simulations to extract parameters requiring optimization, such as absorption bandwidth and absorptivity, and further fine-tune structural parameters until the metasurface structure design is complete. However, the selection of artificial parameters often involves uncertainty, easily falling into local optima and leading to design deviations. Therefore, this approach limits design freedom and makes further performance improvements difficult.

[0003] Combining coding methods with optimization algorithms can effectively improve design freedom and enhance the electromagnetic absorption bandwidth of materials. This approach divides the metasurface structure into an N×N matrix and controls the filling or non-filling of the matrix to achieve the design of the metasurface structure. Although this method significantly improves design freedom, it faces a severe computational efficiency bottleneck: during the algorithm's iterative optimization process, evaluating the electromagnetic response of each transparent absorber structure requires full-wave electromagnetic simulation, which often takes several minutes per simulation. A single loop requires tens of thousands of algorithm iterations, making the enormous time and computational costs prohibitive for designing broadband, high-performance transparent absorbers. Using deep learning models to replace full-wave electromagnetic simulation can theoretically effectively reduce evaluation time and is an effective solution. However, model training requires a large amount of full-wave simulation data, and traditional neural network models have limited prediction accuracy for complex, multi-parameter structures. Therefore, there is an urgent need for multi-parameter neural network training models that can handle small datasets to meet the requirements for broadband performance prediction and optimization of transparent absorbers.

[0004] Existing transparent absorber design methods have the following problems:

[0005] Uniform thin film structures have limited performance characteristics: uniform thin film structures have problems such as narrow absorption bandwidth, large thickness, and low visible light transmittance.

[0006] Manual design has limitations: the selection of manual parameters is often uncertain, which can easily lead to local optima and design deviations. The degree of design freedom is limited, and it is difficult to further improve performance.

[0007] The computational efficiency bottleneck is severe: In the process of algorithm iteration and optimization, the evaluation of the electromagnetic response of each transparent absorber structure requires full-wave electromagnetic simulation. A single structural simulation often takes several minutes. The huge time and computing power costs make the design of broadband high-performance transparent absorbers unbearable.

[0008] Challenges in deep learning training: Model training requires a large amount of full-wave simulation data, and traditional neural network models have limited prediction accuracy for complex multi-parameter structures. Summary of the Invention

[0009] This invention primarily addresses the following problems in the prior art:

[0010] Uniform thin film structures have limited performance characteristics: uniform thin film structures have problems such as narrow absorption bandwidth, large thickness, and low visible light transmittance.

[0011] Manual design has limitations: the selection of manual parameters is often uncertain, which can easily lead to local optima and design deviations. The degree of design freedom is limited, and it is difficult to further improve performance.

[0012] The computational efficiency bottleneck is severe: In the process of algorithm iteration and optimization, the evaluation of the electromagnetic response of each transparent absorber structure requires full-wave electromagnetic simulation. A single structural simulation often takes several minutes. The huge time and computing power costs make the design of broadband high-performance transparent absorbers unbearable.

[0013] Challenges in deep learning training: Model training requires a large amount of full-wave simulation data, and traditional neural network models have limited prediction accuracy for complex multi-parameter structures.

[0014] Therefore, this invention proposes a design method for a multi-parameter broadband transparent absorber based on a deep learning model, characterized by the following steps:

[0015] Step 1: Construct a multi-parameter absorbing unit structure, wherein the absorbing unit includes a top-level metal coding matrix and multiple physical structural parameters;

[0016] Step 2: Randomly generate initial matrix units and obtain the top-level coding pattern based on the transformation of the initial matrix units. Use a data pre-screening algorithm to generate coding matrix samples for the top-level coding pattern, and use full-wave electromagnetic simulation to obtain the absorption rate labels at the corresponding frequencies to construct a multi-physical structure parameter training dataset.

[0017] Step 3: Using the encoding matrix and physical structure parameters as inputs and the electromagnetic response as output, construct a positive prediction model;

[0018] Step 4: Train the positive prediction model using a progressive training strategy to obtain a performance prediction model with multi-parameter input; the progressive training strategy includes:

[0019] Initial training phase: The encoding matrix and some physical structure parameters are set as variables, while the remaining parameters are set as quantitative parameters for training;

[0020] Progressive training phase: Based on the model trained in the previous phase, parameters set to a certain quantity are gradually released as new variables for subsequent training; in the progressive training phase, a preset proportion of existing samples are randomly selected from historical training data and mixed into the current training set for resampling training.

[0021] Step 5: Use the finally trained positive prediction model as a surrogate model for the genetic algorithm to predict the absorption rate, and calculate the relative bandwidth where the absorption rate is higher than the absorption rate threshold within the preset frequency.

[0022] Step 6: Optimize with the goal of maximizing the relative bandwidth where the absorption rate is higher than the threshold within the preset frequency. Determine whether the absorption performance meets the requirements. If the maximum number of evolutions is reached, stop the optimization and output the optimal absorber; if not, generate a new population and repeat step 5.

[0023] Furthermore, the absorbing unit structure comprises, from bottom to top, a bottom metal reflective layer, a middle dielectric layer, and a top metal resonant layer; the metal resonant layer is divided into... A binary matrix, where These are preset constants; in step 1, the physical structural parameters include the air cavity thickness. Unit period and the resistance value of the top metal .

[0024] Furthermore, in step 2, the initial matrix units are randomly generated and have a size of [missing information]. A binary matrix, where The encoding matrix is ​​generated by performing a fourfold rotational symmetry process on the binary initial matrix.

[0025] Furthermore, the data pre-screening algorithm employs the Kennard-Stone algorithm, which calculates the distance matrix between samples and uses the maximum-minimum distance criterion to iteratively select a subset representing global features from the original sample library as input samples for obtaining electromagnetic response labels. Specifically:

[0026] Calculate the L2 norm between every two samples to obtain the inter-sample distance matrix. The calculation formula is:

[0027]

[0028] in, , For any two non-repeating samples, , Number any two samples; based on the calculated distance matrix The Kennard-Stone algorithm iteratively selects two maximum values ​​from the remaining samples to form an initial set S. It then calculates the minimum distance from these remaining samples to the selected samples and adds the sample with the largest minimum distance to S as a new sample. The formula for this iterative selection is as follows:

[0029]

[0030]

[0031] , S is the initial sample number selected from the screening. , The initial set formed, , Number any two remaining non-repeating samples. Number the newly selected samples;

[0032] Pre-screening of training data using the Kennard-Stone algorithm will yield a small dataset representing global features.

[0033] Furthermore, the positive prediction model is a Vision Transformer deep learning model, which includes a graph embedding layer, a scalar fully connected layer, a feature Transformer encoder, and an output head.

[0034] Furthermore, the graph embedding layer maps the input encoding matrix into a preset number of graph tile sequences through convolutional layers.

[0035] Furthermore, the scalar fully connected layer includes multiple independent branch fully connected networks, which map the input physical structure parameters into scalar feature sequences of the same dimension as the tile sequence.

[0036] Furthermore, the feature Transformer encoder is composed of stacked multi-layer encoder modules, and each module employs a multi-head self-attention mechanism to extract global spatial features.

[0037] Furthermore, the output head comprises a multi-layer fully connected network.

[0038] Furthermore, in step 5, the model training process employs the Smooth L1 loss function and uses the AdamW learner for training, while also adopting a dynamic learning rate decay strategy; the model training employs an early stopping strategy to determine the coefficient of determination on the validation set. As a monitoring indicator for the early stopping strategy; the optimization algorithm described in step 6 is a genetic algorithm, whose optimization objective is to maximize the absorption bandwidth within the preset frequency band.

[0039] This invention effectively solves the problem of requiring a large amount of simulation data for deep learning model training by employing a data pre-screening algorithm (Kennard-Stone algorithm) to select small-scale datasets representing global features from the original sample library. In model construction, the Vision Transformer architecture is used to extract global spatial features and achieve multimodal feature fusion. Combined with a progressive training strategy to gradually release variables and incorporate historical sample resampling, this successfully overcomes the challenges of limited prediction accuracy and multi-parameter convergence in complex multi-parameter structures. The proposed method significantly improves the design efficiency of complex multi-parameter transparent absorbers, transforming the optimization process, which originally relied on time-consuming full-wave electromagnetic simulation, into an efficient surrogate model optimization. This not only solves the problem of traditional designs easily getting trapped in local optima but also achieves broadband high-performance absorption with an absorption rate exceeding 90% in the 10–27 GHz frequency band. Attached Figure Description

[0040] Figure 1 The figure shows the training loss and validation loss curves of the deep learning model in this invention for three progressive training iterations, as well as a comparison example of the prediction curve and the electromagnetic simulation curve.

[0041] Figure 2 This is a schematic diagram of the fourfold rotational symmetry generation of the top-layer coding pattern of the coded transparent absorber in this invention, and a schematic diagram of the unit structure of the metal resonant layer-air layer-metal reflective layer.

[0042] Figure 3 The electromagnetic absorption rate spectrum of the optimal transparent absorber structure designed by the present invention using a genetic algorithm combined with a deep learning model in the 0~30GHz frequency band is shown.

[0043] Figure 4 This is a flowchart of the Kennard-Stone algorithm used for pre-screening training data in this invention.

[0044] Figure 5 The present invention provides an overall optimization flowchart of a broadband design method for multi-parameter transparent absorbers based on a deep learning model;

[0045] Figure 6This is a schematic diagram of the network structure of the forward prediction Vision Transformer deep learning model constructed in this invention. Detailed Implementation

[0046] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. The following examples will help those skilled in the art to further understand the present invention, but do not limit the present invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Specific implementation method one:

[0048] This embodiment provides a design method for a multi-parameter broadband transparent absorber based on a deep learning model, including the following steps:

[0049] Step 1: Constructing an absorbing unit structure, the bottom layer of the unit structure is a metal reflective layer, the middle layer is an air cavity, and the top layer is a metal resonant layer. The top metal layer of the absorbing unit structure is divided into... A binary matrix, where 0 represents no metal filling and 1 represents metal filling;

[0050] Step 2: Randomly generate a large number of... A binary matrix of size 1.5 is used as the initial matrix unit, combined with... Figure 2 The top-level encoding matrix is ​​obtained by performing a four-fold rotational symmetry. The Kennard-Stone algorithm is then used to pre-screen the generated encoding matrix, resulting in an encoding matrix with a sample size of 10,000. The pre-screened encoding matrix is ​​then matched with the air cavity thickness of randomly generated unit structures. Unit period and the resistance value of the top metal ,in Select a range of 1~10 mm, including the boundary; Select a range of 5~25 mm, including the boundary; The range of 100~150 Ω, including the boundary, was selected. The data was input into CST for electromagnetic simulation. The output S-parameters were post-processed to obtain the corresponding absorbance. Cubic spline interpolation of the absorbance was performed using 501 sampling points to obtain the training set labels.

[0051] Combination Figure 4 The Kennard-Stone algorithm, a data pre-screening algorithm, is described as follows:

[0052] Calculate the L2 norm between every two samples to obtain the inter-sample distance matrix. The matrix The calculation formula is:

[0053]

[0054] in, , For any two non-repeating samples, , Let any two samples be numbered; based on the calculated distance matrix D, select the two maximum values ​​from it as the initial set S, and calculate the minimum distance from the remaining samples to the selected samples. Select the data with the largest minimum distance as a new sample and add it to S. The Kennard-Stone algorithm's iterative selection formula is expressed as:

[0055]

[0056]

[0057] in, , S is the initial sample number selected from the screening. , The initial set formed, , Number any two remaining non-repeating samples. Number the newly selected samples;

[0058] Pre-screening of training data using the Kennard-Stone algorithm will yield a small dataset representing global features.

[0059] Step 3: Using the encoding matrix, dielectric thickness, cell period, and top-layer metal resistance as training inputs, and the corresponding absorption rate as training outputs, and performing normalization processing respectively, a forward prediction Vision Transformer deep learning model is constructed:

[0060] Combination Figure 6 The forward prediction model Vision Transformer consists of a graph embedding layer, a scalar fully connected layer, a feature Transformer encoder, and an output head.

[0061] The first layer of the graph embedding layer contains 128 elements. The first layer is a convolutional kernel; the second layer is a Flatten layer; and the third layer transposes the feature dimension and the channel dimension to output a sequence of 25 patch tokens.

[0062] The scalar fully connected layer consists of three independent fully connected layers. The first fully connected layer has an input dimension of 1 and an output dimension of 64, and is activated by GELU. The second fully connected layer has an input dimension of 64 and an output dimension of 128, and outputs a total of 3 scalar feature sequences (Scalartokens).

[0063] Subsequently, a 128-dimensional learnable classification token ([CLS] token), a sequence of 25 graph tiles output from the graph embedding layer, and 3 scalar feature sequences are concatenated along the sequence dimension to form an input sequence with a total length of 29, and a learnable positional embedding of the corresponding length is added.

[0064] The Transformer encoder consists of an input layer normalization layer, a 6-layer Transformer encoder module, and an output layer normalization layer. Each encoder module contains an 8-head self-attention mechanism and a feedforward neural network with a hidden layer dimension of 1024. The module uses the GELU activation function and sets a dropout rate of 10%.

[0065] The output header extracts the [CLS] token and inputs it into a 5-layer fully connected network for absorption rate prediction.

[0066] The first layer has an input dimension of 128 and an output dimension of 512, with batch normalization, GELU activation function, and a 10% dropout rate;

[0067] The second layer has an input dimension of 512 and an output dimension of 1024, with batch normalization, GELU activation function, and a 10% dropout rate.

[0068] The third layer has an input dimension of 1024 and an output dimension of 2048, with batch normalization, GELU activation function, and a 10% dropout rate.

[0069] The fourth layer has an input dimension of 2048 and an output dimension of 1024, with batch normalization, GELU activation function, and a 10% dropout rate.

[0070] The fifth layer has an input dimension of 1024 and an output dimension of 501, corresponding to the predicted absorption rate length.

[0071] Step 4: Combining Figure 1 The initial model was trained using the encoding matrix and dielectric thickness as input variables, and the structural period and top-layer metal resistance as quantitative parameters. Subsequently, based on this initial model, the structural period was converted into a variable for secondary training. Finally, based on the secondary training model, the top-layer metal resistance was further incorporated as an input variable for training. Additionally, in each subsequent progressive training iteration, 10% of the samples were randomly selected from the historical training data and mixed into the current training set.

[0072] Define the loss function as the Smooth L1 loss function. When the prediction error is greater than... When using L1 loss, if the prediction error is less than Switching to L2 loss at the time, three training sessions We set the values ​​to 0.85, 0.08, and 0.75 respectively; we used the AdamW learner for training, and adopted a dynamic learning rate decay strategy. The validation set loss was used as the monitoring metric. If the monitoring metric did not decrease for 10 consecutive epochs, the learning rate was decayed to 60% of the current value. The initial learning rate for the first training was set to 5e-4, and the weight decay was set to 1e-4; the initial learning rate for the second training was set to 2e-4, and the weight decay was set to 2e-4; the initial learning rate for the first training was set to 2e-4, and the weight decay was set to 3e-4.

[0073] The model training employs an early stopping strategy, continuously monitoring the validation set determination coefficients during training. If the current value exceeds the historical high, the model parameters are saved; if it exceeds the validation set determination coefficient over 100 consecutive training epochs, the model parameters are saved. If the historical high value is not exceeded, training will be terminated early.

[0074] Step 5: Use the finally trained multi-parameter input positive prediction model Vision Transformer as a genetic algorithm surrogate model to predict the absorption rate, and calculate the relative bandwidth with an absorption rate higher than 90% in the range of 0~30 GHz.

[0075] Step 6: Combining Figure 5 The optimization aims to maximize the relative bandwidth with an absorption rate of over 90% within the 0-30 GHz range. The optimization process is then judged to determine if the absorption performance meets the requirements. If the maximum number of evolutions is reached, the optimization stops and the optimal absorber is output. If not, a new population of individuals is generated, and step 5 is repeated.

[0076] Combination Figure 3 This indicates that the designed structure has good electromagnetic absorption performance, achieving an absorption rate of over 90% within approximately 10–27 GHz.

[0077] The scope of this invention is not limited to the above-described embodiments; a combination of one or more specific embodiments can also achieve the purpose of the invention.

Claims

1. A design method for a multi-parameter broadband transparent absorber based on a deep learning model, characterized in that, Includes the following steps: Step 1: Construct a multi-parameter absorbing unit structure, wherein the absorbing unit includes a top-level metal coding matrix and multiple physical structural parameters; Step 2: Randomly generate initial matrix units and obtain the top-level coding pattern based on the transformation of the initial matrix units. Use a data pre-screening algorithm to generate coding matrix samples for the top-level coding pattern, and use full-wave electromagnetic simulation to obtain the absorption rate labels at the corresponding frequencies to construct a multi-physical structure parameter training dataset. Step 3: Using the encoding matrix and physical structure parameters as inputs and the electromagnetic response as output, construct a positive prediction model; Step 4: Train the positive prediction model using a progressive training strategy to obtain a performance prediction model with multiple parameter inputs; The progressive training strategy includes: Initial training phase: The encoding matrix and some physical structure parameters are set as variables, while the remaining parameters are set as quantitative parameters for training; Progressive training phase: Based on the model trained in the previous phase, parameters set to a certain quantity are gradually released as new variables for subsequent training; in the progressive training phase, a preset proportion of existing samples are randomly selected from historical training data and mixed into the current training set for resampling training. Step 5: Use the finally trained positive prediction model as a surrogate model for the genetic algorithm to predict the absorption rate, and calculate the relative bandwidth where the absorption rate is higher than the absorption rate threshold within the preset frequency. Step 6: Optimize with the goal of maximizing the relative bandwidth where the absorption rate is higher than the threshold within the preset frequency. Determine whether the absorption performance meets the requirements. If the maximum number of evolutions is reached, stop the optimization and output the optimal absorber; if not, generate a new population and repeat step 5.

2. A broadband design method for multi-parameter transparent absorbers based on a deep learning model, as described in claim 1, is characterized in that... The absorbing unit structure, from bottom to top, includes a bottom metal reflective layer, a middle dielectric layer, and a top metal resonant layer; the metal resonant layer is divided into... A binary matrix, where These are preset constants; in step 1, the physical structural parameters include the air cavity thickness. Unit period and the resistance value of the top metal .

3. The broadband design method for multi-parameter transparent absorbers based on a deep learning model according to claim 2, characterized in that, In step 2, the initial matrix unit is randomly generated and has a size of A binary matrix, where The encoding matrix is ​​generated by performing a fourfold rotational symmetry process on the binary initial matrix.

4. The broadband design method for multi-parameter transparent absorbers based on a deep learning model according to claim 1, characterized in that, The data pre-screening algorithm employs the Kennard-Stone algorithm, which calculates the distance matrix between samples and uses the maximum-minimum distance criterion to iteratively select a subset representing global features from the original sample database as input samples for obtaining electromagnetic response labels. Specifically: Calculate the L2 norm between every two samples to obtain the inter-sample distance matrix. The calculation formula is: in, , For any two non-repeating samples, , Number any two samples; based on the calculated distance matrix The Kennard-Stone algorithm iteratively selects two maximum values ​​from the remaining samples to form an initial set S. It then calculates the minimum distance from these remaining samples to the selected samples and adds the sample with the largest minimum distance to S as a new sample. The formula for this iterative selection is as follows: , S is the initial sample number selected from the screening. , The initial set formed, , Number any two remaining non-repeating samples. Number the newly selected samples; Pre-screening of training data using the Kennard-Stone algorithm will yield a small dataset representing global features.

5. The broadband design method for multi-parameter transparent absorbers based on a deep learning model according to claim 1, characterized in that, The positive prediction model is a Vision Transformer deep learning model, which includes a graph embedding layer, a scalar fully connected layer, a feature Transformer encoder, and an output head.

6. The broadband design method for multi-parameter transparent absorbers based on a deep learning model according to claim 5, characterized in that, The graph embedding layer maps the input encoding matrix into a preset number of graph tile sequences through convolutional layers.

7. The broadband design method for multi-parameter transparent absorbers based on a deep learning model according to claim 5, characterized in that, The scalar fully connected layer includes multiple independent branch fully connected networks, which map the input physical structure parameters into scalar feature sequences of the same dimension as the tile sequence.

8. The broadband design method for multi-parameter transparent absorbers based on a deep learning model according to claim 5, characterized in that, The feature Transformer encoder consists of stacked multi-layer encoder modules, with each module employing a multi-head self-attention mechanism to extract global spatial features.

9. The broadband design method for multi-parameter transparent absorbers based on a deep learning model according to claim 5, characterized in that, The output head contains a multi-layer fully connected network.

10. A broadband design method for multi-parameter transparent absorbers based on a deep learning model, as described in claims 1-8, is characterized in that... In the model training process of step 5, the Smooth L1 loss function is used, the AdamW learner is used for training, and a dynamic learning rate decay strategy is adopted. The model training employs an early stopping strategy to optimize the determination coefficients on the validation set. As a monitoring indicator for the early stopping strategy; the optimization algorithm described in step 6 is a genetic algorithm, whose optimization objective is to maximize the absorption bandwidth within the preset frequency band.

Images