Transfer learning-based conformal super-structure surface scattering field rapid prediction method
Through the transfer learning method, the electromagnetic scattering field prediction model is trained, which solves the computational complexity and unit coupling problems in superstructure surface design, and achieves fast and accurate scattering field distribution prediction, improving design efficiency and application prospects.
Patent Information
- Application Number
- CN202411176665.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-11-19
AI Technical Summary
The prior art has problems such as high computational complexity, long simulation time and unconsidered unit coupling in superstructure design, resulting in low computational efficiency of scattering field distribution and inaccurate results.
Using a transfer learning-based method, the sample data set is constructed and data preprocessed. The electromagnetic scattering field prediction model is trained using the first sample data set, the second sample data set and the third sample data set, considering the coupling between units and Shape changes, quickly predict the scattering field distribution of conformal superstructure surface through transfer learning.
It effectively reduces the data set acquisition time and computing resource consumption, establishes a more accurate and fast scattering field prediction model, which can adapt to supersurfaces with different shapes and carrier shapes, and improves design efficiency and practical application prospects.
Smart Images

Figure CN118981967A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of electromagnetic scattering technology, and in particular to a method for quickly predicting the scattering field of a conformal metasurface based on transfer learning. Background Art
[0002] Metasurfaces are generally composed of wavelength and subwavelength metasurface units arranged in a periodic or quasi-periodic manner on a plane. They are two-dimensional plane or curved materials with novel electromagnetic properties that natural media do not have. They are a two-dimensional form of metamaterials that can flexibly control the phase, polarization, amplitude and other properties of electromagnetic waves. They are also flat, compact, low-profile and low-loss. The powerful electromagnetic control capability and ultra-thin characteristics of metasurfaces have made them widely studied in beam focusing / refracting and reflecting, new antenna design, and RCS reduction. Phase is the control phase of the metaunit. For a metasurface, it is the order and distribution of the control phase of each metaunit on it. It is generally represented by a two-dimensional matrix, and the phase value range is between 0 and 360°; the scattering field is the scattering field of the metasurface under plane wave irradiation conditions.
[0003] As a powerful computing tool, deep learning allows computational models to learn data representations and has strong nonlinear fitting and generalization capabilities. It has excellent performance in the fields of classification, translation, and image recognition. In addition to the above mainstream applications of deep learning, it is also used in the fields of optics and electromagnetism, such as the application of deep learning for photonic crystal structure design, reverse design of superstructure units, design of absorbing materials, etc. Although deep learning can replace commercial electromagnetic simulation to a certain extent, it requires a large amount of training data for model training. Using full-wave simulation methods to obtain data sets is not only time-consuming but also computationally expensive. In addition, for metasurfaces with specific structures, once the geometric parameters change slightly, the trained network will not be able to meet the new problems, and the training data set and network model need to be re-established.
[0004] There are still two challenges in the design of metasurfaces based on deep learning. First, due to the higher computational complexity of metasurfaces, it is difficult to spend a lot of time and computing resources to collect training data sets. Using professional electromagnetic simulation software CST Microwave Studio for full-wave simulation, the time required for metasurfaces is much longer than the simulation time of metaunits, and the simulation time will increase significantly as the dimension and size of the metasurface array increase. Considering the time and computing cost issues, it is not feasible to directly use metasurface electromagnetic full-wave simulation to collect data sets. Second, under real scattering conditions, the units on the metasurface are no longer independent, but coupled with each other. The biggest difference between the radiation field of the array calculated based on antenna array theory and the results of full-wave simulation is that the former does not consider the coupling between units, and assumes that the units are independent of each other. It is a scattering field obtained under very ideal conditions, while the latter considers factors such as the coupling and interaction of metaunits in the electromagnetic simulation process, and the simulation results are consistent with the actual electromagnetic scattering situation. Traditional metasurface design methods are generally based on the generalized Snell's law, and the macroscopic arrangement sequence of the corresponding metaunits is designed according to the function of the metasurface. This design generally assumes that the metaunits are independent. But in fact, the potential mutual coupling between adjacent metastructure units will cause certain changes in the electromagnetic properties. Especially when the metastructure surface structure is bent, the changes in the electromagnetic properties will be unpredictable. Summary of the invention
[0005] To this end, the present invention provides a fast prediction method for the scattering field of conformal metasurface based on transfer learning, which solves the problems of inefficient calculation caused by the deviation of scattering field distribution calculated based on theoretical formulas, slow full-wave simulation calculation speed, and the need to re-model and analyze the scattering field distribution calculation of surfaces with different shapes.
[0006] According to the design scheme provided by the present invention, on the one hand, a method for fast prediction of the scattering field of a conformal metasurface based on transfer learning is provided, comprising:
[0007] Constructing a sample data set and performing data preprocessing on the sample data set, the sample data includes a first sample data set, a second sample data set and a third sample data set, the first sample data set is a paired data set of phase and scattered field distribution obtained by calculation according to antenna theory, the second sample data set is a paired data set of phase and scattered field distribution obtained by full-wave simulation using electromagnetic simulation software, and the third sample data set is a paired data set of phase and scattered field distribution obtained by full-wave simulation using electromagnetic simulation software under a conformal shape of a metasurface of a specified target scene;
[0008] The scattering field distribution calculation model is trained by using the sample data set after data preprocessing to obtain an electromagnetic scattering field prediction model, so as to use the electromagnetic scattering field prediction model to predict the scattering field distribution under the conformal shape of the target scene meta-surface;
[0009] Among them, the model training process includes: using the first sample data set to train the scattering field distribution calculation model to obtain the first target model scattering field; using the second sample data set to perform migration training on the first target model to obtain the second target model scattering field; using the third sample data set to perform migration training on the second target model to obtain the electromagnetic scattering field prediction model; the data preprocessing process includes one-hot encoding of the phase distribution and maximum and minimum normalization of the scattering field distribution.
[0010] As a fast prediction method for the scattering field of a conformal metasurface based on transfer learning of the present invention, further, a process scattering field is constructed using the first sample data set, comprising:
[0011] A phase distribution random generation strategy is adopted to obtain multiple sets of phase data, and the size of each set of phase data is N;
[0012] Design conformal metasurfaces based on specified scenarios;
[0013] Based on the phase data and according to the antenna array theory and conformal metasurface, the relationship function between the phase distribution and the scattering field pattern in the planar system is established;
[0014] The phase data is input as a relational function to obtain the corresponding scattered field distribution, thereby forming paired data of phase and scattered field distribution;
[0015] The phase data is converted using one-hot encoding, and the one-hot encoding phase matrix is obtained by mapping the phase data to the Euclidean space, and the scattering field distribution is normalized using the maximum and minimum values of the scattering field distribution data;
[0016] A first sample data set is constructed according to the one-hot encoded phase matrix and the normalized scattering field distribution, so as to perform multi-fold training on the scattering field distribution calculation model based on a mean square error loss function and using the first sample data set.
[0017] As a fast prediction method of the scattering field of a conformal metasurface based on transfer learning of the present invention, further, a conformal metasurface is designed according to a specified scene, comprising:
[0018] The scattering control of UAV wings is selected as the designated scenario, and the conformal metasurface coordinate function is established according to the chord length, relative thickness, relative curvature of the airfoil and the coordinates of the metacell.
[0019] As a fast prediction method of the conformal metasurface scattering field based on transfer learning of the present invention, further, the relationship function between the phase distribution and the scattering field pattern in the plane system is expressed as:
[0020] Where M and N are the sizes of the unit antenna array, both of which are positive integers. Ф(x,y) is the phase distribution function. x, y, and z correspond to the coordinates of the unit antenna array on the metasurface, respectively, and the coordinates are calculated according to the conformal metasurface coordinate function. It is expressed as the scattered field direction parameter θ and the phase distribution parameter Represents the unit scattering field of the unit antenna.
[0021] As a fast prediction method for the scattering field of a conformal metasurface based on transfer learning of the present invention, further, a second sample data set construction process includes:
[0022] The superstructure unit is set as a circular metal patch unit, wherein the circular metal patch unit includes a metal layer, a flexible dielectric layer, a 3D dielectric layer and a metal backplane;
[0023] The phase difference of the circular metal patch unit at the specified frequency and unit radius is obtained through simulation, and the phase corresponding to the unit radius mapping is selected according to the phase difference to construct the target metasurface;
[0024] A phase distribution random generation strategy is adopted to obtain multiple sets of target phase data, each set of phase data is N. The target phase data is used as input, the radius of the circular metal patch unit is adjusted according to the target metasurface, and multiple rounds of full-wave simulation are performed using electromagnetic simulation software to obtain the corresponding scattering field distribution data.
[0025] A second sample data set is constructed according to the target phase data and the corresponding scattering field distribution data in each round of full-wave simulation, and data preprocessing is performed on the second sample data set so as to train the first target model using the second sample data set to obtain a full-wave simulation scattering field prediction model that adapts to the specified shape, and the full-wave simulation scattering field prediction model that adapts to the specified shape is used as the second target model.
[0026] As a fast prediction method of the scattering field of a conformal metasurface based on transfer learning of the present invention, further, the radius of the circular metal patch unit is adjusted according to the target metasurface and multiple rounds of full-wave simulation are performed using electromagnetic simulation software, including:
[0027] Conformal metasurfaces were constructed in an electromagnetic simulation software environment, and periodic boundaries were set in the y-axis direction, and open add space boundaries were set in the x- and z-directions. The metasurfaces were irradiated with linearly polarized plane waves propagating along the -z direction, and the scattered field distribution results were extracted using a scattered field monitor at a specified frequency.
[0028] MATLAB is used to control electromagnetic simulation software for automated simulation. According to the mapping between phase and unit radius in the target metasurface, the unit radius is adjusted using the target phase data to perform multiple rounds of full-wave simulation.
[0029] As a fast prediction method for the scattering field of a conformal metasurface based on transfer learning of the present invention, further, a third sample data set construction process includes:
[0030] The phase data is obtained by adopting a phase distribution random generation strategy; the phase data is used as input, and multiple rounds of full-wave simulations are performed using electromagnetic simulation software under the target scene metasurface after conformal shape adjustment to obtain the corresponding scattering field distribution data;
[0031] A third sample data set is constructed based on the phase data and the corresponding scattered field distribution data in each round of full-wave simulation.
[0032] On the other hand, the present invention also provides a method for fast prediction of conformal metasurface scattering field based on transfer learning, comprising: a sample construction module and a model training module, wherein:
[0033] A sample construction module is used to construct a sample data set and perform data preprocessing on the sample data set, wherein the sample data includes a first sample data set, a second sample data set and a third sample data set, wherein the first sample data set is a paired data set of phase and scattering field distribution obtained by calculation based on antenna theory, the second sample data set is a paired data set of phase and scattering field distribution obtained by full-wave simulation using electromagnetic simulation software, and the third sample data set is a paired data set of phase and scattering field distribution obtained by full-wave simulation using electromagnetic simulation software under a conformal shape of a specified target scene metasurface;
[0034] A model training module is used to train the scattering field distribution calculation model using the sample data set after data preprocessing to obtain an electromagnetic scattering field prediction model, so as to use the electromagnetic scattering field prediction model to predict the scattering field distribution under the conformal shape of the target scene meta-surface;
[0035] Among them, the model training process includes: using the first sample data set to train the scattering field distribution calculation model to obtain the first target model scattering field; using the second sample data set to perform migration training on the first target model to obtain the second target model scattering field; using the third sample data set to perform migration training on the second target model to obtain the electromagnetic scattering field prediction model; the data preprocessing process includes one-hot encoding of the phase distribution and maximum and minimum normalization of the scattering field distribution.
[0036] Beneficial effects of the present invention:
[0037] The present invention utilizes the consistency of theoretical calculation and full-wave simulation in physical mechanism, and uses a large amount of green and economical theoretical training data to replace the time-consuming and complex full-wave simulation electromagnetic scattering field data for model pre-training, which can effectively reduce the time of obtaining data sets and save computing resource consumption. Transfer learning is performed on the basis of the theoretical data pre-training model, and a small amount of full-wave simulation scattering field scattering data set can be used to build an accurate and fast conformal metasurface scattering field prediction model; model training is performed based on the full-wave simulation method, considering the interaction and coupling between units, and the uncertain electromagnetic response caused by conformal bending, etc., which is closer to the actual scattering situation than the scattering field data obtained by theoretical scattering field calculation, so the established model is more realistic and accurate, and the calculation time of the trained network model is much less than the time required for full-wave simulation; the original flat metasurface is expanded to a curved surface, so that the metasurface can be shaped according to the carrier shape, which increases the application scenarios of the metasurface, further promotes the metasurface to practical application, and has a good application prospect. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] Figure 1 It is a schematic diagram of the fast prediction process of the scattering field of the conformal metasurface based on transfer learning in the embodiment;
[0039] Figure 2 This is a comparison of the convergence speed of model training under the conditions of no transfer and transfer learning in the embodiment;
[0040] Figure 3 It is a schematic diagram of the average test loss under different proportions of pre-training model migration and no migration conditions in the embodiment;
[0041] Figure 4 Schematic diagram of the fitting effect of the scattering field prediction model under two randomly selected test examples in the embodiment;
[0042] Figure 5 Schematic diagram of the comparison of model adaptability after the shape of the metasurface is changed in the embodiments. DETAILED DESCRIPTION
[0043] In order to make the purpose, technical solutions and advantages of the present invention clearer and more understandable, the present invention is further described in detail below in conjunction with the accompanying drawings and technical solutions.
[0044] Since the existing metasurface design does not take into account the coupling between units, the scattering field results of the metasurface (especially conformal metasurface) obtained by theoretical calculation of the antenna array are greatly different from the actual scattering situation. Although the simulation using electromagnetic software takes into account the interaction and coupling between units under real scattering conditions, the calculation time is too long, which is not realistic for deep learning networks that require a large number of data sets for training. In the embodiment of this case, an intelligent calculation model of phase distribution to scattering field is constructed based on a deep convolutional neural network. A pre-trained model for scattering field calculation is obtained by training on the theoretical model. Then, for phase distributions of different shapes, the scattering field distribution is quickly and accurately migrated and calculated under the condition of a small amount of full-wave simulation sample data, which is suitable for the automatic inference of scattering field distributions of various shapes, greatly improving the efficiency of phase to scattering field distribution calculation.
[0045] See also Figure 1 As shown, in the embodiment of this case, the process of the method for rapid prediction of the scattering field of a conformal metasurface based on transfer learning can be designed to include the following contents:
[0046] S101, construct a sample data set and perform data preprocessing on the sample data set, the sample data includes a first sample data set, a second sample data set and a third sample data set, the first sample data set is a paired data set of phase and scattering field distribution obtained by antenna theory calculation, the second sample data set is a paired data set of phase and scattering field distribution obtained by full-wave simulation using electromagnetic simulation software, and the third sample data set is a paired data set of phase and scattering field distribution obtained by full-wave simulation using electromagnetic simulation software under a conformal shape of a specified target scene metasurface. The data preprocessing process includes one-hot encoding of the phase distribution and maximum and minimum normalization of the scattering field distribution.
[0047] Specifically, the first sample data set construction process can be designed to include the following contents:
[0048] A phase distribution random generation strategy is adopted to obtain multiple sets of phase data, and the size of each set of phase data is N;
[0049] Design conformal metasurfaces based on specified scenarios;
[0050] Based on the phase data and according to the antenna array theory and conformal metasurface, the relationship function between the phase distribution and the scattering field pattern in the planar system is established;
[0051] The phase data is input as a relational function to obtain the corresponding scattered field distribution, thereby forming paired data of phase and scattered field distribution;
[0052] The phase data is converted using one-hot encoding, and the one-hot encoding phase matrix is obtained by mapping the phase data to the Euclidean space, and the scattering field distribution is normalized using the maximum and minimum values of the scattering field distribution data;
[0053] A first sample data set is constructed according to the one-hot encoded phase matrix and the normalized scattering field distribution, so as to perform multi-fold training on the scattering field distribution calculation model based on a mean square error loss function and using the first sample data set.
[0054] The second sample data set construction process can be designed to include the following contents:
[0055] The superstructure unit is set as a circular metal patch unit, wherein the circular metal patch unit includes a metal layer, a flexible dielectric layer, a 3D dielectric layer and a metal backplane;
[0056] The phase difference of the circular metal patch unit at the specified frequency and unit radius is obtained through simulation, and the phase corresponding to the unit radius mapping is selected according to the phase difference to construct the target metasurface;
[0057] A phase distribution random generation strategy is adopted to obtain multiple sets of target phase data, each set of phase data is N. The target phase data is used as input, the radius of the circular metal patch unit is adjusted according to the target metasurface, and multiple rounds of full-wave simulation are performed using electromagnetic simulation software to obtain the corresponding scattering field distribution data.
[0058] A second sample data set is constructed according to the target phase data and the corresponding scattering field distribution data in each round of full-wave simulation, and the first target model is trained using the second sample data set to obtain a full-wave simulation scattering field prediction model that adapts to the specified shape, and the full-wave simulation scattering field prediction model that adapts to the specified shape is used as the second target model.
[0059] Among them, in multiple rounds of full-wave simulation, a conformal metasurface was constructed in the electromagnetic simulation software environment, and periodic boundaries were set in the y-axis direction, and open add space boundaries were set in the x and z directions; a linearly polarized plane wave propagating along the -z direction was used to irradiate the metasurface, and the scattering field distribution results were extracted using a scattering field monitor at a specified frequency; MATLAB was used to control the electromagnetic simulation software for automated simulation, and according to the mapping between the phase and the unit radius in the target metasurface, the unit radius was adjusted using the target phase data for multiple rounds of full-wave simulation.
[0060] The third sample data set construction process can be designed to include the following contents:
[0061] Adjust the conformal shape of the metasurface under the target scene; use the phase distribution random generation strategy to obtain phase data; use the phase data as input, use electromagnetic simulation software to perform multiple rounds of full-wave simulation under the metasurface of the target scene after the conformal shape is adjusted, and obtain the corresponding scattering field distribution data;
[0062] A training sample data set is constructed based on the phase data and the corresponding scattered field distribution data in each round of full-wave simulation.
[0063] S102. Using the sample data set after data preprocessing to train the scattering field distribution calculation model, obtain an electromagnetic scattering field prediction model, and use the electromagnetic scattering field prediction model to predict the scattering field distribution under the conformal shape of the target scene meta-surface.
[0064] The model training process includes: using the first sample data set to train the scattering field distribution calculation model to obtain the first target model scattering field; using the second sample data set to perform migration training on the first target model to obtain the second target model scattering field; using the third sample data set to perform migration training on the second target model to obtain the electromagnetic scattering field prediction model.
[0065] Furthermore, based on the above method, an embodiment of the present invention also provides a fast prediction system for conformal metasurface scattering field based on transfer learning, comprising: a sample construction module and a model training module, wherein:
[0066] A sample construction module is used to construct a sample data set and perform data preprocessing on the sample data set, wherein the sample data includes a first sample data set, a second sample data set and a third sample data set, wherein the first sample data set is a paired data set of phase and scattering field distribution obtained by calculation based on antenna theory, the second sample data set is a paired data set of phase and scattering field distribution obtained by full-wave simulation using electromagnetic simulation software, and the third sample data set is a paired data set of phase and scattering field distribution obtained by full-wave simulation using electromagnetic simulation software under a conformal shape of a specified target scene metasurface;
[0067] A model training module is used to train the scattering field distribution calculation model using the sample data set after data preprocessing to obtain an electromagnetic scattering field prediction model, so as to use the electromagnetic scattering field prediction model to predict the scattering field distribution under the conformal shape of the target scene meta-surface;
[0068] Among them, the model training process includes: using the first sample data set to train the scattering field distribution calculation model to obtain the first target model scattering field; using the second sample data set to perform migration training on the first target model to obtain the second target model scattering field; using the third sample data set to perform migration training on the second target model to obtain the electromagnetic scattering field prediction model; the data preprocessing process includes one-hot encoding of the phase distribution and maximum and minimum normalization of the scattering field distribution.
[0069] In order to verify the effectiveness of this solution, the following is a further explanation based on experimental data:
[0070] The steps of the scattered field prediction algorithm during the experiment can be summarized as follows:
[0071] Step 1: Random generation of phase distribution. A random generation strategy is used to obtain n groups of one-dimensional phase data of size N.
[0072] Step 2: Solve the conformal unit coordinates. Taking the scattering control of the drone wing as the research scenario, the Zhukovsky profile line is selected to design the conformal metasurface, and the metaunit coordinates satisfy the following formula:
[0073]
[0074] Among them, parameter l is the chord length of the airfoil; δ represents the ratio of the maximum thickness of the wing to the chord length, that is, the relative thickness; ε represents the ratio of the maximum distance from the wing mid-curve to the chord to the chord length, that is, the relative curvature, and the parameter values are l = 300mm, δ = 0.15, ε = 0.015. x, z are the coordinates of the superstructure unit, the units are arranged at equal distances along the curve, and the units are arranged at equal intervals along the straight line in the y-axis direction, and the coordinates are integer multiples of the unit period p.
[0075] Step 3: Calculation of theoretical scattered field data. Using phase data as input, according to antenna array theory, the relationship between phase distribution and scattered field pattern in a planar system is as follows:
[0076]
[0077] Among them, M, N are the unit array sizes, M and N are both positive integers, Ф(x,y) is the phase distribution function, x, y, z correspond to the coordinates of the unit on the metasurface, which are related to the geometric shape of the conformal metasurface and can be calculated based on the conformal geometric surface function, k is the wave number, k = 2π / λ. Among them, the phase distribution Ф is discretized, and the values are 0, 90°, 180°, and 270°, respectively. The corresponding scattering field distribution is calculated to form paired data of phase and scattering field distribution, and a theoretical training data set for scattering field distribution calculation is constructed. In this patent, the phase distribution has a gradient in only one dimension. Let the size of this dimension be N=24. Since there is no phase change in the other dimension, the phase distribution can be represented by a one-dimensional vector. The phase distribution selects the one-dimensional vector phase data generated in step 1.
[0078] Step 4: Phase encoding. Based on one-hot encoding, the data set The phase distribution is converted, the discrete phase distribution is mapped to the Euclidean space, and the original one-dimensional vector is expanded into a two-dimensional matrix. The matrix dimension is N×4, where N corresponds to the phase dimension size, and 4 is the number of discrete phases. In this patent, the phase takes four values: 0°, 90°, 180°, and 270°. For example, for the i-th (i≤N) phase, if its value is 90°, the i-th row and the second column element is 1, and the remaining column elements of the i-th row are 0. For example, for the case of N=5, if the one-dimensional phase distribution is (0, 90, 270, 180, 270), the phase matrix converted to one-hot encoding is According to the data generated in the one-hot encoding step 1 above, the phase distribution is converted from 0 to 1 to obtain n groups of one-hot encoding phase distributions with a size of 24×4.
[0079] Step 5: Normalization of scattered field data. Based on formula (3), the scattered field distribution data is normalized. Here, the maximum and minimum rule is used to complete the normalization operation.
[0080]
[0081] Step 6: Represent the preprocessed phase data and scattered field data as X and Y respectively to form a large-scale theoretical training data set
[0082] Step 7: Construct a scattered field distribution calculation model The model contains 2 convolution blocks and 5 fully connected blocks. The convolution block consists of a 2D convolution layer, a regularization layer, and an activation function. The number of output channels of the convolution layer in the two convolution blocks is 32 and 64 respectively, and the activation function uses the LeakyReLU function; the fully connected block consists of a fully connected layer and an activation function, and the number of neurons in the five fully connected layers is 4096, 2048, 1024, 512, and 181 respectively.
[0083] Step 8: Calculate the scattered field distribution model based on the theoretical training data set Multi-fold training is performed, the learning rate of training is set to 0.0002, the batch size is 128, and the loss function adopts the mean square error loss function, as shown in formula (4).
[0084]
[0085] Step 9: After multiple rounds of training, the scattered field distribution calculation model is obtained and store it.
[0086] Step 10: Design of meta-unit. In theoretical calculations, the specific meta-unit structure and the interaction between units are not considered, and the phase value can be set directly. In the actual preparation process of the meta-surface, the design of the specific unit structure needs to be considered. This patent selects a circular metal patch unit. The unit contains a total of four layers, of which the first layer is a metal layer with a radius of r and a thickness of 0.036 mm. The second layer is a flexible dielectric layer with a thickness of 0.2 mm and a dielectric constant of 2.65. The third layer is a 3D dielectric layer with a thickness of 1.8 mm and a dielectric constant of 2.7. The fourth layer is a metal backplane. The unit structure is shown in the figure. The unit period p = 9 mm. Through simulation, it is found that the unit is at a frequency of 12 GHz. The phase difference is 90° when the metal patch radius r = 1, 3, 3.46 and 4.4 mm. Therefore, the above four structures are selected to map four discretized phases for meta-surface construction.
[0087] Step 11: Use the same random strategy as step 1 to generate k groups of phase data (k≤0.2n), use this as input, and use CST electromagnetic simulation software to perform full-wave simulation to obtain scattered field distribution data. Since full-wave simulation takes into account the interaction and coupling between units, the scattered field data obtained using electromagnetic simulation software is more realistic.
[0088] Step 12: Full-wave simulation scattering field data acquisition. Electromagnetic calculations were performed using CST and Matlab co-simulation. First, the CST simulation conditions were set. A conformal metasurface was constructed in the CST simulation environment according to formula (1). The y-axis direction was set as a periodic boundary, and the x and z directions were set as open add space boundaries. The linearly polarized plane wave propagating along the -z direction was used to irradiate the metasurface. A scattering field monitor at a frequency of 12 GHz was set to extract the scattering field results. Secondly, CST was controlled by MATLAB for automated simulation to adjust the geometric radius size of the circular metal patch. The radius of the circular metal patch along the x direction was selected according to the mapping between the phase and the geometric radius in step 10 and the randomly generated phase in step 11. After the construction was completed, the full-wave simulation was started. Then, after the scattering field results were calculated, the 181 scattering field distribution data corresponding to the phase distribution were extracted and the scattering field results were stored. Next, according to the phase data in step 11, the geometric size of the circular metal patch is automatically changed, and a new round of full-wave simulation is performed. The scattering field results corresponding to the new round of phase data are extracted and stored, and this cycle is repeated until k rounds of full-wave simulation calculations are completed. The obtained k groups of phase and scattering field results are preprocessed according to steps 4 and 5 to construct a full-wave simulation electromagnetic scattering training data set.
[0089] Step 13: Load the pre-trained model
[0090] Step 14: Full-wave simulation scattered field prediction model Structure and pre-trained theoretical model in step 7 The structure is similar, and the loss function is set to formula (4). Based on the full-wave simulation electromagnetic model Initialize the parameters and set the model Migrate to Relying on data sets right Train to obtain a full-wave simulation scattering field prediction model that adapts to the current shape
[0091] Step 15: Storing the Real Electromagnetic Scattering Model Parameters.
[0092] Step 16: To further explore the effectiveness and reliability of small sample transfer learning, change the conformal shape of the conformal metasurface and perform model training and analysis verification. Change the relative thickness δ and relative curvature ε in formula (1) in step 2 to set δ = 0.2 and ε = 0.03. Construct the corresponding full-wave simulation electromagnetic model data set under the new conformal shape according to steps 11 and 12.
[0093] Step 17: For the new shape, collect a small amount of phase data and its corresponding scattered field distribution data, process the data using the preprocessing operations of steps 4 and 5, and construct a lightweight training data set corresponding to the new shape
[0094] Step 18: Use the full-wave simulation electromagnetic model in step 15 As a pre-trained model, the scattering field calculation model of the current shape is based on this Initialize the parameters and use them in the training data set Transfer learning is performed on the scattering field prediction model of the secondary migration to obtain
[0095] Step 19: One-hot encode the given phase data to obtain the phase input x.
[0096] Step 20: For the phase input x, rely on the electromagnetic scattering prediction model Automatically infer its scattered field distribution y.
[0097] The experimental contents and experimental results based on the above algorithm design are as follows:
[0098] Experiment 1: Analysis of the convergence speed and accuracy of the model under different data scales.
[0099] The theoretical electromagnetic scattering data set for phase and scattered field distribution is 57,521 data items, and the full-wave simulation data set is 13,041 data items. There is no redundant data in the data set. The theoretical model for scattering field distribution calculation is trained based on the theoretical data set. For a given shape S1 (δ = 0.15, ε = 0.015), based on the full-wave simulation electromagnetic scattering data set, the scattering field distribution calculation model constructed by direct training is compared with the scattering field distribution calculation model constructed by direct training. The model training convergence speed and test accuracy of the scattering field distribution calculation model in two cases of transfer learning as a pre-trained model.
[0100] Figure 2 (a) to (c) are the training process of the scattering field distribution calculation model under the condition of no migration, and the training data set accounts for 0.1, 0.2, and 0.3 of the total data set respectively; (d) to (f) are the training process of the scattering field distribution calculation model under the condition of transfer learning, and the training data set accounts for 0.1, 0.2, and 0.3 of the total data set respectively. Figure 2 As shown in the figure, under the condition of no transfer, the model needs to be trained for 30 to 40 rounds to converge asymptotically, while under the condition of transfer learning, the model only needs to be trained for about 20 rounds to converge asymptotically, with a faster convergence speed. Moreover, under the condition of a small number of samples, the transfer learning model can converge to a better loss value range, and the loss value reduction is optimized by 1 order of magnitude under the conditions of 1174 and 2647 samples compared with the condition of no transfer, and the loss value reduction is optimized by 1 order of magnitude under the condition of no transfer in the validation set.
[0101] Figure 3 The red line in the figure is the average loss of the model on the test data set under the transfer learning based on the pre-trained model, and the blue line in the figure is the average loss of the model on the test data set without transfer. The horizontal axis is the proportion of the training data set to the total data set, and the value range is 0.1 to 0.9. Figure 3 As shown in the figure, under the condition that the total amount of the data set remains unchanged, when the proportion of the training data set is less than 30%, the model based on transfer learning can achieve a better average loss value, and as the proportion of the training data set increases, the average loss value shows a significant downward trend, which proves the performance advantage of transfer learning under the condition of few samples of electromagnetic data.
[0102] Experiment 2: Analysis of the effect of the scattering field prediction model under different conformal structures.
[0103] For a given shape S1 (δ = 0.15, ε = 0.015), under a given phase condition, relying on the transfer learning model The corresponding scattered field distribution is obtained, and the distribution difference between the calculated results after migration and the full-wave simulation electromagnetic measurement values is observed and analyzed. Figure 4 As shown in the figure, two test cases are randomly selected and the model prediction results are compared with the actual values. It can be seen that the obtained scattering field is consistent with its true value peak distribution, especially the extreme value points in the scattering field distribution have a high degree of consistency, which shows that the model The prediction effect is good. The goodness of fit is used to evaluate the prediction effect of the network. The calculation formula for the goodness of fit is as follows:
[0104]
[0105] R 2 The closer it is to 1, the better the fitting effect is. According to formula (5), Figure 4 The goodness of fit of the network model predictions in (a) and (b) is 94.26% and 90.66%, respectively. Figure 4 In the comparison results of scattered field distribution under different phase conditions, the blue line is the true value and the orange line is the network predicted value.
[0106] For a given shape S2 (δ = 0.2, ε = 0.03), in the model Based on the second migration, the model was constructed For comparison, a model constructed directly from a small amount of shape S2 full-wave simulation data and a model based on shape S1 theoretical data were also studied. The simulation results of the model after a migration are as follows: Figure 5 As shown in the figure, (a) the non-transferred learning network model built using a small amount of S2 full-wave simulation data, and (b) the S1 theoretical prediction model Based on the S2 full-wave simulation data, a migration network model is performed. (c) Scattering prediction model is transferred from S2 full-wave simulation (first migration) Based on this, the network model is migrated again (second migration) using S2 full-wave simulation data
[0107] Figure 5(a) is the scattering field result predicted by the model trained with a small amount of S2 full-wave simulation data without transfer learning, (b) is the scattering field result predicted by the model after one-time transfer learning, using S2 full-wave simulation data for migration based on the S1 theoretical prediction model, and (c) is the scattering field result predicted by the model after two-time transfer learning, using S1 full-wave simulation data for migration based on the S1 theoretical prediction model, and then using S2 full-wave simulation data for migration again on this basis. The blue line in the figure is the real full-wave simulation data. The goodness of fit of (a), (b), and (c) are 33.46%, 71.45%, and 87.41%, respectively, indicating that the two-time transfer model is more effective.
[0108] like Figure 5 As shown in (a) in the figure, when the shape structure being analyzed changes, the electromagnetic computational model trained under the condition of a small amount of scattering field distribution data set for the current shape has multiple peaks of the scattering field distribution shifted laterally, and the fitting degree is poor. Figure 5 As shown in (b) in the figure, after a migration based on the theoretical model, the electromagnetic calculation model obtained based on the migration of the theoretical model overcomes the deficiency of the current small size of shape training data, and the distribution of the peak value of the scattered field distribution is corrected, but there are still large deviations and fluctuations at some points; Figure 5 As shown in (c) in the figure, based on the model After the migration, the final scattering field distribution calculation model adapted to the electromagnetic characteristics of the new shape, and the accuracy of the scattering field distribution was greatly improved under the condition of a small training data set. Its average error was reduced by about 73.9% compared with the theoretical model, and its peak value calculation error was improved by more than 16.8% compared with the original shape electromagnetic model before migration.
[0109] Simultaneous use and The time required to predict the scattering field of a metasurface is only milliseconds, which is much shorter than the full-wave simulation time. Therefore, the transfer learning network can support the agile design of conformal metasurfaces.
[0110] The above experimental results show that, for phase distributions of different shapes, this scheme can quickly and accurately transfer and calculate the scattering field distribution under the conditions of a small amount of full-wave simulation sample data, adapt to the automatic inference of scattering field distributions of various shapes, and greatly improve the efficiency of phase to scattering field distribution calculation. The average error of scattering field distribution calculation is reduced by about 73.9% compared with the theoretical calculation model. After transfer learning, the prediction accuracy of the peak value of the scattering field distribution of the model to adapt to the new shape is improved by more than 16.8%. It has good application prospects in the agile design of conformal metasurfaces, electromagnetic wave control and propagation.
[0111] Unless otherwise specifically stated, the relative steps, numerical expressions and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.
[0112] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims
1. A fast prediction method for the scattering field of conformal metasurface based on transfer learning, characterized in that: Include: Constructing a sample data set and performing data preprocessing on the sample data set, the sample data includes a first sample data set, a second sample data set and a third sample data set, the first sample data set is a paired data set of phase and scattered field distribution obtained by calculation according to antenna theory, the second sample data set is a paired data set of phase and scattered field distribution obtained by full-wave simulation using electromagnetic simulation software, and the third sample data set is a paired data set of phase and scattered field distribution obtained by full-wave simulation using electromagnetic simulation software under a conformal shape of a metasurface of a specified target scene; The scattering field distribution calculation model is trained by using the sample data set after data preprocessing to obtain an electromagnetic scattering field prediction model, so as to use the electromagnetic scattering field prediction model to predict the scattering field distribution under the conformal shape of the target scene meta-surface; Among them, the model training process includes: using the first sample data set to train the scattering field distribution calculation model to obtain the first target model scattering field; using the second sample data set to perform migration training on the first target model to obtain the second target model scattering field; using the third sample data set to perform migration training on the second target model to obtain the electromagnetic scattering field prediction model; the data preprocessing process includes one-hot encoding of the phase distribution and maximum and minimum normalization of the scattering field distribution.
2. The method for rapid prediction of conformal metasurface scattering fields based on transfer learning according to claim 1, characterized in that: Construct the process scattered field using the first sample data set, including: A phase distribution random generation strategy is adopted to obtain multiple sets of phase data, and the size of each set of phase data is N; Design conformal metasurfaces based on specified scenarios; Based on the phase data and according to the antenna array theory and conformal metasurface, the relationship function between the phase distribution and the scattering field pattern in the planar system is established; The phase data is input as a relational function to obtain the corresponding scattered field distribution and form paired data of phase and scattered field; The phase data is converted using one-hot encoding, and the one-hot encoding phase matrix is obtained by mapping the phase data to the Euclidean space, and the scattering field distribution is normalized using the maximum and minimum values of the scattering field distribution data; A first sample data set is constructed according to the one-hot encoded phase matrix and the normalized scattering field distribution, so as to perform multi-fold training on the scattering field distribution calculation model based on a mean square error loss function and using the first sample data set.
3. The method for rapid prediction of conformal metasurface scattering fields based on transfer learning according to claim 2, characterized in that: Design conformal metasurfaces based on specified scenarios, including: The scattering control of UAV wings is selected as the designated scenario, and the conformal metasurface coordinate function is established according to the chord length, relative thickness, relative curvature of the airfoil and the coordinates of the metacell.
4. The method for rapid prediction of conformal metasurface scattering fields based on transfer learning according to claim 3, characterized in that: The relationship function between phase distribution and scattered field pattern in a planar system is expressed as: Where M and N are the sizes of the unit antenna array, both of which are positive integers. Ф(x,y) is the phase distribution function. x, y, and z correspond to the coordinates of the unit antenna array on the metasurface, respectively, and the coordinates are calculated according to the conformal metasurface coordinate function. is the scattered field direction parameter θ and the phase distribution parameter Represents the unit scattering field of the unit antenna.
5. The method for rapid prediction of conformal metasurface scattering fields based on transfer learning according to claim 1, characterized in that: The second sample data set construction process includes: The superstructure unit is set as a circular metal patch unit, wherein the circular metal patch unit includes a metal layer, a flexible dielectric layer, a 3D dielectric layer and a metal backplane; The phase difference of the circular metal patch unit at the specified frequency and unit radius is obtained through simulation, and the phase corresponding to the unit radius mapping is selected according to the phase difference to construct the target metasurface; A phase distribution random generation strategy is adopted to obtain multiple sets of target phase data, and the size of each set of phase data is N; Taking the target phase data as input, the radius of the circular metal patch unit is adjusted according to the target metasurface, and multiple rounds of full-wave simulation are performed using electromagnetic simulation software to obtain the corresponding scattering field distribution data. A second sample data set is constructed according to the target phase data and the corresponding scattering field distribution data in each round of full-wave simulation, and data preprocessing is performed on the second sample data set so as to train the first target model using the second sample data set to obtain a full-wave simulation scattering field prediction model that adapts to the specified shape, and the full-wave simulation scattering field prediction model that adapts to the specified shape is used as the second target model.
6. The method for rapid prediction of conformal metasurface scattering fields based on transfer learning according to claim 5, characterized in that: The radius of the circular metal patch unit is adjusted according to the target metasurface and multiple rounds of full-wave simulations are performed using electromagnetic simulation software, including: Conformal metasurfaces were constructed in an electromagnetic simulation software environment, and periodic boundaries were set in the y-axis direction, and open add space boundaries were set in the x- and z-directions. The metasurfaces were irradiated with linearly polarized plane waves propagating along the -z direction, and the scattered field distribution results were extracted using a scattered field monitor at a specified frequency. MATLAB is used to control electromagnetic simulation software for automated simulation. According to the mapping between phase and unit radius in the target metasurface, the unit radius is adjusted using the target phase data to perform multiple rounds of full-wave simulation.
7. The method for rapid prediction of conformal metasurface scattering fields based on transfer learning according to claim 1, characterized in that: The third sample data set construction process includes: The phase data is obtained by adopting a phase distribution random generation strategy; the phase data is used as input, and multiple rounds of full-wave simulations are performed using electromagnetic simulation software under the target scene metasurface after conformal shape adjustment to obtain the corresponding scattering field distribution data; A third sample data set is constructed based on the phase data and the corresponding scattered field distribution data in each round of full-wave simulation.
8. A fast prediction system for conformal metasurface scattering fields based on transfer learning, characterized in that: Contains: sample construction module and model training module, among which, A sample construction module is used to construct a sample data set and perform data preprocessing on the sample data set, wherein the sample data includes a first sample data set, a second sample data set and a third sample data set, wherein the first sample data set is a paired data set of phase and scattered field distribution obtained by calculation based on antenna theory, the second sample data set is a paired data set of phase and scattered field distribution obtained by full-wave simulation using electromagnetic simulation software, and the third sample data set is a paired data set of phase and scattered field distribution obtained by full-wave simulation using electromagnetic simulation software under a conformal shape of a specified target scene metasurface; A model training module is used to train the scattering field distribution calculation model using the sample data set after data preprocessing to obtain an electromagnetic scattering field prediction model, so as to use the electromagnetic scattering field prediction model to predict the scattering field distribution under the conformal shape of the target scene meta-surface; Among them, the model training process includes: using the first sample data set to train the scattering field distribution calculation model to obtain the first target model scattering field; using the second sample data set to train the first target model to obtain the second target model scattering field; using the third sample data set to train the second target model to obtain the electromagnetic scattering field prediction model; the data preprocessing process includes one-hot encoding of the phase distribution and maximum and minimum normalization of the scattering field distribution.
Citation Information
Cited By
Prediction method for unit scattered field of variable-curvature frequency selective surface
CN120068649A
