A deep learning-based electromagnetic simulation multi-scale region segmentation method

Abstract

The application discloses a kind of electromagnetic simulation multiscale region segmentation methods based on deep learning, constructs and trains deep learning model, for identifying key area and non-key area in electromagnetic simulation region;Training process uses labeled data set and uses cross-validation method to optimize model performance;Key area is simulated using full-wave solution method, and the field distribution of key area is obtained;Non-key area is simulated using high-frequency approximation method, and the electromagnetic field generated after the reflection of incident wave by object surface is simulated;Dynamic boundary condition is generated at the boundary of key area and non-key area, to ensure the continuity and consistency of field value.The application realizes the optimized configuration of computing resources in large-scale electromagnetic simulation scene by the combination of deep learning and hybrid solution strategy, focuses the limited computing power on high-precision demand area, and significantly improves the simulation efficiency and overall performance.

Description

TECHNICAL FIELD

[0001] The present invention relates to the field of electromagnetic simulation, and particularly relates to a multi-scale region segmentation method for electromagnetic simulation based on deep learning, which is used to optimize the computational efficiency and accuracy in the process of electromagnetic simulation. BACKGROUND

[0002] Large-scale electromagnetic simulation involves complex electromagnetic wave propagation, scattering and radiation problems, and is widely used in the fields of antenna design, radar cross section (RCS) analysis, electromagnetic compatibility (EMC) testing, wireless communication coverage, etc. With the increase of simulation scale, traditional small-scale simulation methods have encountered bottlenecks in terms of computing time and resource demand, so various advanced techniques have emerged: full-wave solution, high-frequency approximation, hybrid method, parallel computing and high-performance computing. Although existing electromagnetic simulation techniques have made significant progress, there are still several major problems in dealing with large-scale electromagnetic simulation:

[0003] 1) High computational complexity and resource demand: In large-scale problems, the computational complexity of full-wave solution increases sharply with frequency, geometric size and grid number. Especially in high-frequency, large-size simulation, the sparse matrix generated by full-wave method is very large, and solving these matrices requires a large amount of memory and long computing time.

[0004] 2) Trade-off between accuracy and efficiency: Although high-frequency approximation methods can improve efficiency in large-size, far-field problems, they lack accuracy in complex geometric structures or near-field regions because they ignore some details in the electromagnetic field. Hybrid methods can balance accuracy and efficiency, but they require high requirements for region division and boundary condition handling. Manually setting these region divisions and boundary conditions not only consumes time, but also introduces errors.

[0005] 3) Complex region division and boundary condition handling: In hybrid methods, different regions need to use different numerical methods, and manually dividing the regions of interest and non-interest may not be accurate enough, leading to inaccurate simulation results or waste of computing resources.

[0006] 4) Lack of simulation automation and intelligence: In the current simulation process, many steps (such as region division, mesh generation, boundary condition setting, etc.) require human intervention, and the degree of automation is low. In large-scale simulation processes, these manual operations are time-consuming and prone to errors. SUMMARY

[0007] Invention purposes: In view of the above technical deficiencies, the purpose of the present application is to provide an electromagnetic simulation multi-scale region segmentation method based on deep learning, to solve the problems of high complexity of large-scale electromagnetic simulation calculation, difficulty in balancing precision and efficiency, complexity of region division and boundary condition processing, and lack of simulation automation and intelligence in the prior art.

[0008] Technical scheme: The electromagnetic simulation multi-scale region segmentation method based on deep learning comprises the following steps:

[0009] (1) Construct and train a deep learning model to identify key regions and non-key regions in the electromagnetic simulation region; the training process uses a labeled data set and uses a cross-validation method to optimize the model performance;

[0010] (2) The key region is simulated by using a full-wave solution method to obtain the field distribution of the key region;

[0011] (3) The non-key region is simulated by using a high-frequency approximation method to simulate the electromagnetic field generated by the reflection of the incident wave on the surface of the object;

[0012] (4) Dynamic boundary conditions are generated at the boundary between the key region and the non-key region to ensure the continuity and consistency of the field value.

[0013] Further, the process of identifying the key region and the non-key region in the electromagnetic simulation region in step (1) is as follows:

[0014] The deep learning model is a convolutional neural network, which extracts features from image data and electromagnetic field distribution data of the simulation region top view; wherein the feature extraction layer includes an edge detection module and a gradient analysis module; the edge detection module extracts the contour information of the region, and the convolution kernel is initialized as a Sobel operator:

[0015]

[0016] After convolution, the gradient amplitude is calculated:

[0017]

[0018] The gradient analysis module is used to capture the change pattern of the electromagnetic field intensity, and the gradient features are extracted by using a convolution kernel of a Gaussian Laplace operator:

[0019]

[0020] In the pooling layer, the feature map is down-sampled to reduce the data dimension:

[0021] f pool (x)=max{x1,x2,...,x n}

[0022] The extracted features are input into the full connection network at the full connection layer, and the probabilities of the key region and the non-key region are output at the classification layer, so as to realize the boundary segmentation of the key region and the non-key region.

[0023] Further, the training process of step (1) adopts a labeled data set and uses a cross-validation method to optimize the model performance, and the implementation process is as follows:

[0024] The cross-entropy loss function is used to measure the gap between the predicted value and the true label:

[0025]

[0026] Where y i is the actual label, is the model output; divide the data into k subsets, and use k-1 of them to train and 1 to validate each time; loop k times to calculate the average validation performance.

[0027] Further, step (2) simulates the key region using the full-wave solution method, which is implemented by using the finite element method or the finite difference time domain method.

[0028] Further, the implementation process of step (2) is as follows:

[0029] The key region is further divided into a finite number of small sub-regions by using the finite element method, and the solution of the local equation is used to approximate the solution of the global equation in each small region; the Maxwell equation is converted into a weak form, that is:

[0030]

[0031] The entire calculation region is divided into small finite elements by using meshing; it is assumed that the electric field and the magnetic field are polynomial functions on each element, and the problem of each small element is solved one by one and assembled into a global equation, and finally the field distribution of the entire region is obtained.

[0032] Further, step (3) simulates the non-key region using the high-frequency approximation method, which is implemented by using physical optics or ray tracing to solve the electromagnetic field.

[0033] Further, the implementation process of step (3) is as follows:

[0034] For electromagnetic waves on the surface, it is assumed that the propagation direction of the reflected wave and the incident wave is only related to the surface normal and the incident angle; according to the basic formula of physical optics:

[0035]

[0036] The electromagnetic field generated after the incident wave is reflected by the surface of the object is simulated by calculating the non-key region.

[0037] Further, the step (4) is implemented as follows:

[0038] According to the key area and the non-key area, the boundary area field value is weighted and averaged using a weighted polynomial interpolation; assuming that at the boundary, the field value of the key area is f1(x), and the field value of the non-key area is f2(x), the field value f(x) at the boundary is expressed by a weighted polynomial, which is in the form of:

[0039] f(x)=w(x)f1(x)+(1-w(x))f2(x)

[0040] Wherein, w(x) is a weight function, and the weight function is designed using quadratic interpolation:

[0041]

[0042] Wherein, x1 and x2 are the positions on both sides of the boundary, and the weight function will realize a continuous transition at the boundary with the change of position.

[0043] Advantages: Compared with the prior art, the advantages of the present application are: 1. Intelligent region division: with the help of deep learning technology, the transition from manual division to automatic generation is realized, and reasonable and multi-scale region segmentation scheme can be automatically generated according to the actual distribution characteristics of electromagnetic field; for the fine calculation area identified, high-precision full-wave solution method is used for detailed simulation; for the brief calculation area, high-efficiency high-frequency approximation method is applied to reduce the calculation complexity; 2. Accurate boundary condition generation: using the prediction ability of deep learning, the change rule of field value between regions is accurately captured, and accurate and reasonable boundary conditions are automatically generated, which effectively reduces the error in the boundary processing process and ensures the seamless connection between different calculation regions; 3. Efficient computing resource allocation: in large-scale electromagnetic simulation scenarios, through the combination of deep learning and hybrid solving strategy, the optimal allocation of computing resources is realized, and the limited computing power is focused on high-precision demand areas, which significantly improves the simulation efficiency and overall performance. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is the flowchart of the present application;

[0045] Figure 2 is the schematic diagram of electromagnetic simulation region segmentation. DETAILED DESCRIPTION

[0046] The technical solutions of the present application are further described below in conjunction with the drawings.

[0047] As shown in Figure 1 , the present application proposes a multi-scale region segmentation method for electromagnetic simulation based on deep learning, and the specific implementation process is as follows:

[0048] Step 1: Construct and train a deep learning model for identifying key and non-key regions in the electromagnetic simulation area. The training process uses a labeled dataset and the cross-validation method to optimize model performance.

[0049] The deep learning model is a convolutional neural network that extracts features from image data and electromagnetic field distribution data of the simulation area top view. The feature extraction layer includes an edge detection module and a gradient analysis module. The edge detection module extracts the contour information of the region, and its convolution kernel is initialized as the Sobel operator:

[0050]

[0051] After convolution, calculate the gradient amplitude:

[0052]

[0053] The gradient analysis module captures the change pattern of the electromagnetic field intensity by extracting gradient features with a Gaussian Laplace operator convolution kernel:

[0054]

[0055] In the pooling layer, the feature map is down-sampled to reduce the data dimension:

[0056] f pool (x)=max{x1,x2,...,x n}

[0057] In the fully connected layer, the extracted features are input into the fully connected network, and the probability of key regions (1) and non-key regions (0) is output in the classification layer. In the process of training and optimizing the model, the cross-entropy loss function is used to measure the gap between the predicted value and the true label:

[0058]

[0059] where y i is the actual label, is the model output. Divide the data into k subsets, and use k-1 for training and 1 for validation each time. Loop training k times and calculate the average validation performance.

[0060] Step 2: Use full-wave solving method for simulation in key regions, including finite element method (FEM) and finite difference time domain method (FDTD).

[0061] Full-wave solution method can comprehensively consider the propagation, scattering and interaction of electromagnetic waves in space, and has high accuracy. The key region divided in step 1 is further divided into a finite number of small sub-regions, and the solution of the local equation is used to approximate the solution of the global equation in each small region. Convert Maxwell's equation into weak form, that is:

[0062]

[0063] Using meshing, the entire calculation region is divided into small finite elements, assuming that the electric field and magnetic field are polynomial functions on each element, by solving each small element problem one by one and assembling into global equations, the field distribution of the entire region is finally obtained.

[0064] Step 3: Use high-frequency approximation method to simulate non-key regions, including physical optics (PO) and ray tracing (RT).

[0065] In the high-frequency limit, the reflection and refraction of electromagnetic fields can be described by simple geometric optics laws. For electromagnetic waves on the surface, it is assumed that the propagation direction of the reflected wave and the incident wave is only related to the surface normal and the incident angle. According to the basic formula of physical optics:

[0066]

[0067] According to the calculation of the non-key region divided in step 1, the electromagnetic field generated by the reflection of the incident wave on the surface of the object is simulated.

[0068] Step 4: Generate dynamic boundary conditions at the boundary between the key region and the non-key region to ensure the continuity and consistency of the field values.

[0069] According to the boundary region segmented by the convolutional neural network in step 1, use weighted polynomial interpolation to perform weighted averaging on the field values of the boundary region. Assuming that the field value of the key region at the boundary is f1(x), and the field value of the non-key region is f2(x), the field value f(x) at the boundary is expressed by weighted polynomial:

[0070] f(x) = w(x)f1(x) + (1-w(x))f2(x)

[0071] Where w(x) is the weight function, and quadratic interpolation is used to design the weight function:

[0072]

[0073] Where x1 and x2 are the positions on both sides of the boundary, and the weight function will realize a continuous transition at the boundary with the change of position.

[0074] In this embodiment, the electromagnetic simulation region is determined, for example, in the simulation of the impact of a wind farm on a radar. The geometry, material properties of the wind turbines in the wind farm, and the operating frequency of the radar need to be determined. After the electromagnetic simulation region is determined, the simulation region is manually labeled to mark the key regions and non-key regions. As shown in Figure 2 The simulation region is distributed with many wind turbines, and the areas where the wind turbines are densely distributed are the key regions in the simulation, while the areas where the wind turbines are sparsely distributed or where there are no wind turbines are the non-key regions. In the key regions, the electromagnetic field is finely simulated using the full-wave method, and in the non-key regions, the electromagnetic field is simply simulated using the high-frequency approximation method, thereby generating high-quality training data for the deep learning model to provide input and labels.

[0075] The UNet segmentation model in the convolutional neural network is selected to process the spatial data of the electromagnetic field distribution, and the input data is the electromagnetic field distribution data and the image data of the simulation region. The output data is the region division mask (1 represents the full-wave solving region, and 0 represents the high-frequency approximation region). The cross-entropy loss is used to measure the difference between the predicted value and the true label, and the model parameters are optimized according to the value.

[0076] In the full-wave solving region, the finite element method (FEM) and the time-domain finite difference method (FDTD) are used to perform high-precision electromagnetic field solving, and in the high-frequency approximation region, the physical optics (PO) and the ray tracing (RT) are used to quickly solve the electromagnetic field. At the boundary between the full-wave solving region and the high-frequency approximation region, the field values are coupled to ensure that the field solutions of different solving regions can seamlessly connect and ensure physical accuracy.

[0077] The field value distribution of the full-wave solving region and the high-frequency approximation region is compared to check the continuity of the field value at the boundary to ensure that the boundary conditions are correctly applied. The simulation results are compared with experimental data or theoretical solutions (such as analytical solutions and approximate solutions) to ensure simulation accuracy. The sources of simulation errors are identified, and the grid density and time step problems are analyzed to optimize and adjust the errors. If the boundary conditions or region division are not accurate enough, the feedback mechanism of the deep learning model is used to re-adjust the region division or boundary condition settings to further improve the accuracy of the simulation results.

[0078] The above describes the method of the present application in detail, but it is obvious that the specific implementation form of the present application is not limited to this. Various obvious changes made by those skilled in the art without departing from the spirit and scope of the claims of the present application are within the scope of protection of the present application.

Claims

1. A deep learning-based electromagnetic simulation multi-scale region segmentation method, characterized in that, The method comprises the following steps: (1) constructing and training a deep learning model for identifying key areas and non-key areas in an electromagnetic simulation region; The training process uses a labeled data set and uses a cross-validation method to optimize model performance; (2) simulating the key areas using a full-wave solution method to obtain the field distribution of the key areas; (3) simulating the non-key areas using a high-frequency approximation method to simulate the electromagnetic field generated after the incident wave is reflected by the object surface; (4) generating dynamic boundary conditions at the boundary between the key areas and the non-key areas to ensure the continuity and consistency of the field values; The implementation process of step (4) is as follows: According to the key areas and non-key areas, use weighted polynomial interpolation to perform weighted averaging on the field values of the boundary areas; at the boundary, the field value of the key area is f1(x), and the field value of the non-key area is f2(x), use weighted polynomial to express the field value f(x) at the boundary, which is in the form of: f(x) = w(x)f1(x) + (1-w(x))f2(x) Where w(x) is the weight function, and the weight function is designed using quadratic interpolation: Where x1 and x2 are the positions on both sides of the boundary, and the weight function will realize continuous transition at the boundary as the position changes.

2. The method of claim 1, wherein, The implementation process of step (1) for identifying key areas and non-key areas in an electromagnetic simulation region is as follows: The deep learning model is a convolutional neural network that extracts features from image data and electromagnetic field distribution data of the simulation region top view; Where the feature extraction layer includes an edge detection module and a gradient analysis module; the edge detection module extracts the contour information of the region, and the convolution kernel is initialized as a Sobel operator: After convolution, calculate the gradient amplitude: The gradient analysis module is used to capture the change pattern of the electromagnetic field intensity, and the gradient features are extracted by using a convolution kernel of a Gaussian Laplace operator: In the pooling layer, the feature map is down-sampled to reduce the data dimension: f pool (x) = max{x1,x2,...,x n} In the fully connected layer, the extracted features are input into the fully connected network, and the probability of key areas and non-key areas is output in the classification layer, so as to realize the boundary segmentation of key areas and non-key areas.

3. The method of claim 1, wherein, The implementation process of step (1) for training process using labeled data set and using cross-validation method to optimize model performance is as follows: Use cross-entropy loss function to measure the difference between predicted value and true label: where y i is the actual label, is the model output; split the data into k subsets, train on k-1 of them and validate on the remaining one each time; loop k times and compute the average validation performance.

4. The method of claim 1, wherein, Step (2) uses the full-wave solution method to simulate the key areas, which uses the finite element method or the finite difference time domain method.

5. The method of claim 1, wherein, The implementation process of step (2) is as follows: Use the finite element method to further divide the key areas into a finite number of small sub-areas, and solve the local equation in each small area to approximate the solution of the global equation; convert Maxwell equation into weak form, that is: Divide the entire calculation region into small finite elements by meshing; the electric field and magnetic field are polynomial functions on each element, and the field distribution of the entire region is obtained by solving the problem of each small element one by one and assembling it into a global equation.

6. The method of claim 1, wherein, Step (3) uses the high-frequency approximation method to simulate the non-key areas, which uses physical optics or ray tracing to solve the electromagnetic field.

7. The method of claim 1, wherein, The implementation process of step (3) is as follows: For the electromagnetic wave on the surface, the propagation direction of the reflected wave and the incident wave is only related to the surface normal and the incident angle; according to the basic formula of physical optics: The non-critical area is calculated, and the electromagnetic field generated after the incident wave is reflected by the surface of the object is simulated.

Images