Sequential multi-physics machine learning assisted low drag antenna design method and system

Abstract

The application provides a machine learning assisted antenna structure layer parameterized topology design and optimization method, and based on this, a low wind resistance antenna design method assisted by machine learning of sequential multi-physical fields is provided.In the antenna design method, first, the structure of the antenna metal ground is changed through the parameterized topology design method.Second, the topology structure of the metal ground is optimized to meet the design index of low wind resistance; in the optimization process, a machine learning algorithm is used to establish a proxy model to predict the wind resistance value of the antenna.Finally, the electromagnetic performance of the antenna meeting the wind resistance design target is optimized, and the electromagnetic performance is improved by changing the topology structure of the dipole and the structure parameters of the antenna; in the electromagnetic performance optimization process, a machine learning algorithm is also used to establish a proxy model to predict the bandwidth, isolation and gain of the antenna, so as to accelerate the optimization design process.The application can reduce the design space and reduce the optimization difficulty.

Description

Technical Field

[0001] This invention belongs to the field of antenna design technology, specifically relating to machine learning-assisted low-drag antenna design, which can be used in the field of multiphysics research and design of antennas. Background Technology

[0002] Low wind resistance is a crucial requirement in the design and research of 5G ultra-large-scale base station antennas. On the one hand, to meet the requirements for gain, bandwidth, and other aspects, the number of antennas increases dramatically, making the site selection and installation of base station antennas more difficult. Lower wind resistance can effectively reduce the difficulty of installing and setting up base station antennas. On the other hand, while meeting the physical requirement of low wind resistance, base station antenna design also needs to meet the original electromagnetic performance requirements.

[0003] For dual-polarized crossed dipole antennas widely used in base station antennas, low-drag design requires not only miniaturization of the radiating layer but also topology optimization of the metallic ground to obtain innovative metallic ground structures that facilitate airflow. To quantitatively analyze and compare different metallic ground topologies, simulation analysis using a fluid dynamics simulation solver is necessary. Furthermore, to meet electromagnetic performance requirements, simply optimizing the metallic ground topology is insufficient; optimization of the radiator's topology and structural parameters is also required. Due to the large number of design parameters, relying solely on fluid dynamics and full-wave simulations is insufficient to obtain optimal solutions.

[0004] Over the past decade, machine learning methods have been widely applied to the optimization design of antennas, achieving good results. However, the design objectives of low-drag antennas involve two different physical fields, resulting in numerous optimization goals and a sharp increase in the number of design parameters, creating a huge design space. This leads to an exponential increase in the number of samples, unsustainable sampling time, and the non-negligible training time of the surrogate model. Therefore, the key issue is how to reduce the design space and accelerate the convergence speed of optimization, thereby speeding up the design process of low-drag antennas. Summary of the Invention

[0005] Purpose of the invention: This invention proposes a machine learning-assisted parametric topology design and optimization method for antenna structure layers, and based on this parametric topology design and optimization method, proposes a sequential multiphysics machine learning-assisted low-drag antenna design method, which can reduce the huge design space under the multiphysics design requirements while meeting antenna performance requirements, and accelerate the antenna design process.

[0006] Technical Solution: To achieve the above-mentioned objectives, this invention first provides a machine learning-assisted method for parameterized topology design and optimization of antenna structure layers, comprising the following steps:

[0007] Set an initial mesh image and set the initial material properties of the mesh;

[0008] Set grid selection design parameters for selecting the location and number of grid points, and material property design parameters for changing the material properties of the selected grid points and their adjacent grid points;

[0009] Setting complex image generation design parameters is used to generate more complex mesh images based on mesh images with modified material properties. These parameters include those that determine the combination method of mesh image transformation, the combination order of the original mesh image and the transformed mesh image, and the number of combinations. For image rotation transformation, parameters that determine the rotation angle are also included.

[0010] For more complex mesh images generated based on the above design parameters, the contour lines between meshes with different material properties are extracted, and the topology corresponding to a set of design parameters is obtained by modeling based on the contour lines.

[0011] For the defined antenna performance optimization objective, a surrogate model is established using machine learning methods to assist in performance optimization, thereby obtaining the topology design parameters of the structural layer that meet the design objective.

[0012] Preferably, the parametric topology design and optimization method uses machine learning methods to train the surrogate model, and obtains the training dataset of the surrogate model by sampling within the range of values ​​of all design parameters and performing simulation.

[0013] Specifically, when the parametric topology design and optimization method is used for antenna metallic ground parametric topology design and optimization, an antenna wind resistance performance optimization target is defined, and a proxy model based on the wind resistance value obtained from the projected area of ​​the metallic ground is established using machine learning methods to assist in performance optimization, thereby obtaining metallic ground topology design parameters that meet the design target.

[0014] Specifically, when the parametric topology design and optimization method is used for the parametric topology design and optimization of antenna dipoles, an antenna electromagnetic performance optimization target is defined, and a surrogate model is established using machine learning methods to obtain S-parameters and gain based on design parameters and frequency to assist in performance optimization, thereby obtaining dipole topology design parameters that meet the design target.

[0015] Furthermore, to simultaneously meet the requirements of wind resistance and electromagnetic performance, after optimizing the wind resistance performance of the antenna's metallic ground, the electromagnetic performance of the antenna as a whole is optimized by combining the metallic ground with the mesh support structure. The design parameters include parameters that determine the number of grids per row / column of the mesh support structure, the size of the region in the radiating layer that is not subject to topology optimization, the size of the gap between the design area and the edge of the dipole arm, the diameter of the radiating layer, the size of the support structure around the coaxial feed line, the top angle of the dipole arm, and the distance between the radiating layer and the metallic ground.

[0016] Based on the above-mentioned parametric topology design and optimization method, this invention provides a sequential multiphysics machine learning-assisted low-drag antenna design method, comprising the following steps:

[0017] Based on the parametric topology design and optimization method, the topology of the antenna metallic ground and the dipole are changed by mesh selection design parameters, material property design parameters and complex image generation design parameters, respectively, to optimize the antenna's wind resistance and electromagnetic performance; among which, the antenna's structural parameters are increased for electromagnetic performance optimization.

[0018] We construct a sequential design process for antenna drag performance optimization and electromagnetic performance optimization. We use machine learning methods to establish a surrogate model with low computational complexity in the antenna drag performance optimization process to accelerate the design process and obtain the topology design parameters of the metallic ground that meet the drag design target.

[0019] After assembling a complete cross-dipole antenna by combining a metal ground plane, a radiating layer, and a coaxial feed line that meet the wind resistance design target, machine learning methods are then used to accelerate the electromagnetic performance optimization process of the antenna. This yields a dipole topology and some structural parameters of the antenna that meet the requirements for bandwidth, isolation, and gain, ultimately resulting in a dual-polarized cross-dipole antenna that satisfies both low wind resistance design requirements and electromagnetic performance requirements.

[0020] Furthermore, the antenna's wind resistance performance is optimized by taking into account prior knowledge and the wind resistance calculation formula.

[0021] F w =C f A ref q p

[0022] Where F w Represents wind resistance, C f It is the drag coefficient, A ref It is the size of the projected area, q pIt is dynamic pressure and positively correlated with the square of wind speed; utilizing the correspondence between the design parameters of the metallic ground and the projected area of ​​the metallic ground, a low-computational-complexity surrogate model U is established through machine learning algorithms to establish the relationship between the projected area and the wind resistance. P This improves the accuracy of wind resistance prediction.

[0023] Furthermore, the design parameters for electromagnetic performance optimization include both the design parameters of the dipole topology and some structural parameters of the antenna; the electromagnetic performance optimization of the antenna combines dipole topology optimization and antenna structural parameter optimization; during the electromagnetic performance optimization process of the antenna, a surrogate model with low computational complexity is established using machine learning methods to replace time-consuming full-wave simulation, thereby accelerating the antenna design process.

[0024] Furthermore, in the sequential optimization design process, the output of antenna drag performance optimization is the input of electromagnetic performance optimization. The optimal topology design parameters of the metal ground are obtained based on the drag performance optimization, and the metal ground is modeled in the full-wave simulation software. This metal ground, together with the radiating layer and coaxial feeder, forms an antenna model for the next stage of electromagnetic performance optimization. The output of antenna electromagnetic performance optimization affects whether the drag performance optimization process is restarted. To accelerate the optimization design of low drag antennas, antennas that still cannot meet the electromagnetic performance optimization target after reaching the upper limit of the optimization iteration number will be discarded, and drag performance optimization will be restarted to obtain a new metal ground topology.

[0025] The present invention also provides a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the steps of the above-mentioned parameterized topology design and optimization method, or the sequential multiphysics machine learning-assisted low-drag dipole antenna design method.

[0026] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention proposes a low-complexity parametric topology design and optimization method, which meets the design requirements while effectively reducing the size of the design space; (2) The present invention can obtain a metal ground topology that is conducive to air flow and thus reduces wind resistance, which cannot be designed by experience and intuition; (3) The present invention combines wind resistance performance optimization with electromagnetic performance optimization, which expands the research and design scope of antenna multi-physics field; (4) The present invention uses a sequential optimization design process to decompose the multi-objective optimization problem under multi-physics field, reduce the number of design parameters, greatly reduce the design space, and reduce the difficulty of optimization. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the parametric topology design and optimization method according to an embodiment of the present invention.

[0028] Figure 2 This is a schematic diagram of the topology design process for the antenna metallic ground under a set of given design parameters in an embodiment of the present invention.

[0029] Figure 3 This is a flowchart illustrating the method for optimizing the design of a low-drag dual-polarized cross dipole antenna in an embodiment of the present invention.

[0030] Figure 4 This is a side view of the antenna model obtained after modeling based on a set of design parameters in an embodiment of the present invention.

[0031] Figure 5 This is a top view of the metal ground obtained according to a set of design parameters in an embodiment of the present invention.

[0032] Figure 6 This is a top view of the adjusted antenna radiating layer in an embodiment of the present invention.

[0033] Figure 7 This is a top view of the metal ground obtained by optimizing wind resistance performance in an embodiment of the present invention.

[0034] Figure 8 This is a top view of the radiation layer obtained by electromagnetic performance optimization in an embodiment of the present invention.

[0035] Figure 9 The figure shows the full-wave simulation results of the S-parameters and gain of the optimized low-drag dual-polarized cross dipole antenna in this embodiment of the invention. Detailed Implementation

[0036] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0037] like Figure 1As shown in the figure, this invention discloses a machine learning-assisted parametric topology design and optimization method for antenna structure layers. The method first sets an initial mesh image and initial material properties for the mesh; then, it sets mesh selection design parameters for selecting the location and number of mesh points, and material property design parameters for changing the material properties of the selected mesh points and their adjacent mesh points; and it sets complex image generation design parameters to generate more complex mesh images based on the mesh image with changed material properties, including parameters determining the combination method of mesh image transformation, parameters determining the combination order of the original mesh image and the transformed mesh image, and parameters determining the number of combinations; for image rotation transformation, it also includes parameters determining the rotation angle; for the more complex mesh image generated based on the above design parameters, it extracts the contour lines between meshes with different material properties, and models the topology structure corresponding to a set of design parameters based on the contour lines; finally, for the defined antenna performance optimization objective, it uses machine learning methods to establish a surrogate model to assist performance optimization, obtaining the topology design parameters of the structure layer that meet the design objective.

[0038] For example, when designing a specific object (such as a metallic ground or a radiating layer), the parametric topology design method first sets a size of N. p ×N p The initial mesh image, with all meshes having the default initial material property of metal; setting design parameters. (Where q0 is a positive even number, q1 = q0 × q0) Select q1 specific grid points, let q2 = q0 / 2, and set the parameters... The parameter determines the number of rows of grid points to be selected. Determine the number of columns of grid points to be selected; set design parameters. To change material properties, each selected grid point is surrounded by 3, 5, or 8 nearest neighbor grid points, depending on the parameters. Select q1 adjacent grid points according to the parameters. Adjust the selected q1 grid points, q1 corresponding adjacent grid points, and the unparalleled grid points. Select the material properties of the remaining adjacent points; after completing the initial material property changes, for a size of N... p ×N p After flipping or rotating the original grid image, three new grid images are obtained. These new grid images are then combined in a specific order to generate a more complex new image. Parameters are set... The four transformation combinations for the grid image are determined by: first, the original image undergoes a flip or rotation transformation before being combined into a new image; second, whether the subsequent transformation is the same as the previous transformation; and setting parameters. Determine the order in which the original image and the transformed image are combined; set parameters. The number of times images are combined determines the size of the final generated mesh image; setting parameters... The rotation angle for image rotation transformation is determined; using the above method, a more complex mesh image can be obtained, the contour lines between meshes with different material properties can be extracted, and modeling can be performed based on the contour lines to obtain the corresponding topology under a set of design parameters. Examples of the value ranges of the design parameters are shown in Table 1.

[0039] Table 1. Range of values ​​for parametric topology design parameters

[0040]

[0041] in It is an integer. It is a real number.

[0042] Specifically, for the parametric topology design and optimization of low-drag antennas with metallic ground, an antenna drag performance optimization objective can be defined. Machine learning methods can be used to establish a surrogate model that obtains the drag value based on the projected area of ​​the metallic ground to assist performance optimization, resulting in metallic ground topology design parameters that meet the design objective. Similarly, for the parametric topology design and optimization of high-performance antenna dipoles, an antenna electromagnetic performance optimization objective can be defined. Machine learning methods can be used to establish a surrogate model that obtains the S-parameters and gain based on the design parameters and frequency to assist performance optimization, resulting in dipole topology design parameters that meet the design objective.

[0043] The application of the above parametric topology optimization design will be explained in detail below with a specific example of a low-drag dual-polarized crossed dipole antenna design.

[0044] This invention discloses a sequential multiphysics machine learning-assisted design method for low-drag dual-polarized crossed dipole antennas. This method first establishes a parameterized topology optimization design method using 12 design parameters x1, x2, ..., x... 12 The topology of the antenna's metallic ground is changed, and the range of parameter values ​​is shown in Table 2.

[0045] Table 2 Design parameters for metallic ground topology

[0046] <![CDATA[x1,x2,...,x6]]> 1 8 <![CDATA[x7]]> 0 1 <![CDATA[x8]]> 1 8 <![CDATA[x9]]> 1 4 <![CDATA[x 10 ]]> 1 24 <![CDATA[x 11 ]]> 3 6 <![CDATA[x 12 ]]> 0 90 / / /

[0047] Where x1, x2, ..., x 11 x is an integer 12 Let x1 = 8, x2 = 6, x3 = 7, x4 = 6, x5 = 7, x6 = 3, x7 = 1, x8 = 2, x9 = 3, x... 10 =1,x 11 =3,x 12 =32.3; The topology design process for metallic ground is as follows: Figure 2As shown, the specific steps are as follows:

[0048] (T1) Change the material properties of the initial mesh image: Set an 8×8 initial mesh image, with the default property of each mesh being metal. Select 9 mesh points p1, p2, ..., p9 according to parameters x1, x2, ..., x6, where x1, x2, x3 determine the number of rows of these 9 mesh points, and x4, x5, x6 determine the number of columns; each mesh point has 3, 5, or 8 nearest neighbor meshes around it, and select the adjacent mesh points corresponding to p1, p2, ..., p9 according to parameter x8. Change the material properties of these 9 mesh points, the 9 selected adjacent mesh points, and the remaining adjacent mesh points not selected by parameter x8 according to parameter x7: if x7 = 0, the property of mesh points p1, p2, ..., p9 and their selected adjacent mesh points is changed from metal to air, while the properties of other mesh points remain unchanged; if x7 = 1, the property adjustment is reversed.

[0049] (T2) Complex Image Generation: After completing the initial attribute changes, to further increase the possibility of topological changes, more complex mesh images are needed. Three new mesh images are obtained by flipping or rotating the mesh image obtained in (T1). The original mesh image and the new mesh images are then combined in a certain order to generate a more complex new image. The parameter x9 determines the four transformation combination methods for the mesh images: whether the original image is first flipped or rotated before being combined into a new image, and whether the later transformation is the same as the previous transformation. Figure 2 This corresponds to the case where, when x9 = 3, a flipping transformation is performed first, followed by recombination, and the subsequent transformation differs from the previous one. The parameter x... 10 This determines the order in which the original image and the transformed image are combined. Parameter x 11 The number of image combinations determines the size of the final generated mesh image. Parameter x 12 It determines the angle of rotation when the image undergoes a rotation transformation.

[0050] (T3) Contour Extraction: After obtaining a more complex mesh image, the contour lines between the metal and the air are extracted, and the coordinate information of the contour lines is stored in different matrices. The external contour of the metal ground and the position and shape information of the internal ventilation holes are all contained in different contour lines.

[0051] (T4) Modeling: Based on the coordinate information in the contour matrix obtained in the previous step, modeling is performed in HFSS to obtain a new type of metallic ground. After combining it with the radiating layer and coaxial feed line, an antenna model for wind resistance performance optimization can be obtained.

[0052] Based on the above parameterized topology design method, the flowchart of the machine learning-assisted low-drag dual-polarized crossed dipole antenna design method described in this embodiment of the invention is as follows: Figure 3 As shown, the specific optimization design process is as follows:

[0053] (S101) Initialization settings for drag performance optimization: Define the design parameters for drag performance optimization as [x1, x2, ..., x 12 The optimization objective for wind resistance is defined as t. wl .

[0054] (S102) Sampling is performed to obtain the training set for optimizing the wind resistance design target: the optimization parameters x1, x2, ..., x in Table 2 are used. 12 Within the optimization interval, the initial sample size N is set. a =100, using Latin Hypercube Sampling (LHS), and then using the fluid dynamics software Fluent for simulation, the wind resistance values ​​of different antenna models with different metallic ground topologies but the same radiating layer are obtained.

[0055] (S103) Establish or update the surrogate model used to optimize wind resistance performance: from Figure 2 As can be seen, a set of design parameters generates a complex mesh image for metal ground topology design. The projected area s of the metal ground can be obtained by calculating the area of ​​the metal mesh in the mesh image. s is related to x1, x2, ..., x... 12 There is a mapping relationship between them. The projected area of ​​the metal ground is calculated in MATLAB and used as the input parameter for machine learning. The wind resistance value obtained from the simulation in step (S102) is used as the output parameter. A surrogate model U for the wind resistance value is trained using a machine learning algorithm based on Gaussian process regression. P U will be updated online as more data is added to the training set. P .

[0056] (S104) Antenna wind resistance performance optimization and verification: Utilize a global optimization algorithm (such as a genetic algorithm) to optimize the low-computational-complexity surrogate model U established in step (S103). P To optimize wind resistance, the fitness function F is optimized. WL for:

[0057]

[0058] Where s is the projected area of ​​the metal ground, determined according to design parameters x1, x2, ..., x 12 Calculations show that It is the proxy model U P The predicted wind resistance value, It is the uncertainty of the prediction, ω aThis is the lower confidence bound (LCB), where ω is set to ω. a =1. The goal of optimization is to minimize F. WL This achieves the optimization goal of wind resistance performance. The actual wind resistance value is obtained through simulation in Fluent. Termination condition 1 is met: the wind resistance value is less than the set target value t. wl Or reach the number of wind resistance optimization iterations i a If the upper limit is not met, proceed to the next step (S2). If termination condition 1 is not met, add the optimized parameters and the actual wind resistance value to the training set, return to step (S103), and update the surrogate model U. P We will continue to optimize it.

[0059] (S2) Antenna Modeling: Based on the optimal design parameters obtained in step (S104), a metallic ground model is performed. Together with the radiating layer and coaxial feed line, this model forms the antenna for the next stage of electromagnetic performance optimization. The side view of the antenna is shown below. Figure 4 As shown. To increase the physical structural stability of the metal ground, especially to ensure the structural stability around the coaxial feeder, the modeled metal ground 1-1 is combined with the mesh support structure 1-2, as follows. Figure 5 As shown. To minimize the projected area of ​​the radiating layer and the size of the individually packaged radome, the dielectric substrate of the radiating layer was changed from a traditional square to a circle. Figure 6 The structure of the radiating layer is shown in detail. The next stage of electromagnetic performance optimization will optimize the new design region 4 on the dipole and some structural parameters. It is important to note that to facilitate subsequent assembly using studs, nuts, etc., there should be no metallic radiating structures within region 7. The topology optimization of the dipole differs from that of the metallic ground in two aspects:

[0060] 1) In order to reduce the number of electromagnetic optimization design parameters, the size of the initial mesh image was reduced from 8×8 to 4×4, reducing the number of parameters used to select the mesh; only the flip transformation was used in the image transformation.

[0061] 2) Complex images were changed from squares to concentric arcs, making them more suitable for the new design area.

[0062] Fifteen design parameters for optimizing the electromagnetic performance of the antenna are defined as [z1,z2,...,z8,c,r1,r2,d1,d2,h1,h2]. Here, parameters z1,z2,...,z4 determine the position of the selected grid, z5 determines the attribute change of the selected grid, z6 determines the position of the adjacent selected grid, z7 determines the order of image combination, z8 determines the number of image combinations, parameter c is the number of grids per row / column in the mesh support structure 1-2, parameter r1 determines the size of the region in the radiating layer that is not topologically optimized, parameter r2 determines the size of the gap between design region 4 and the edge of dipole arm 5, parameter d1 is the diameter of the radiating layer, parameter d2 determines the size of the support structure around the coaxial feed line 2, parameter h1 determines the size of the apex angle α of the dipole arm, and parameter h2 determines the distance between the radiating layer 3 and the metal ground 1. The value ranges of the above 15 electromagnetic performance optimization parameters are shown in Table 3.

[0063] Table 3 Design parameters for antenna electromagnetic performance optimization

[0064] <![CDATA[z1,z2,z3,z4]]> 1 4 <![CDATA[r1]]> 3.85 7 <![CDATA[z5]]> 0 1 <![CDATA[r2]]> 0.5 2.5 <![CDATA[z6]]> 1 8 <![CDATA[d1]]> 24.1 29.2 <![CDATA[z7]]> 1 24 <![CDATA[d2]]> 6.5 7 <![CDATA[z8]]> 3 6 <![CDATA[h1]]> 8.6 10.9 c 1 4 <![CDATA[h2]]> 18.5 20.5

[0065] Where z1, z2, ..., z8, c are integers, and r1, r2, d1, d2, h1, h2 are real numbers.

[0066] (S301) Initialization settings for electromagnetic performance optimization: Define the design parameters for electromagnetic performance optimization as [z1,z2,...,z8,c,r1,r2,d1,d2,h1,h2], and define the objective of electromagnetic performance optimization as t. bw1 , t bw2 , t iso , t gain , respectively corresponding to the S of the antenna 11 S 22 S 21 And the design specifications for gain.

[0067] (S302) Sampling to obtain a training set for optimizing electromagnetic performance design objectives: Within the optimization interval of the optimization parameters z1, z2, ..., z8, c, r1, r2, d1, d2, h1, h2 shown in Table 3, set the initial sample size N. e =100, using Latin hypercube sampling, and then obtaining the antenna's S-parameters and gain results through full-wave simulation.

[0068] (S303) Establish or update the surrogate model for electromagnetic performance optimization: Using the antenna design parameters z1, z2, ..., z8, c, r1, r2, d1, d2, h1, h2 and frequency as input parameters for machine learning, and the S-parameters and gain as output parameters, train the model using a machine learning algorithm based on Gaussian process regression to obtain the S-parameters. 11 S 22 S 21 And the proxy model U of gain bw1 U bw2 U iso and U gain The proxy model is updated online as the training set changes.

[0069] (S304) Electromagnetic performance optimization and verification of the antenna: Utilize a global optimization algorithm (such as a genetic algorithm) to optimize the low-computational-complexity surrogate model U established in step (S303). bw1 U bw2 U iso and U gain Electromagnetic performance optimization is performed, and the fitness function F is optimized. EM for:

[0070] F EM =max(ω1g1,ω2g2,ω3g3,ω4g4)

[0071] Where ω1, ω2, ω3, and ω4 are penalty values, and g1, g2, g3, and g4 are respectively S 11 S 22 S 21 The sum and gain are functions, specifically expressed as follows:

[0072]

[0073]

[0074]

[0075]

[0076] in, The S-parameters and gain results obtained from the surrogate model prediction It is the uncertainty of the prediction, ω m It is an LCB value equal to 1. The purpose of electromagnetic performance optimization is to make the fitness function F EM Minimum. Verification is performed using full-wave simulation software. If termination condition 2 is met: both S-parameters and gain have reached the optimization target, then the optimal antenna design parameters are output. If termination condition 2 is not met and the iteration count i has not been reached... emIf the upper limit is reached, the optimized parameters and the actual full-wave simulation results are added to the training set, and the process returns to step (S303), the surrogate model is updated, and optimization continues; if termination condition 2 is not met and the iteration count i is reached... em If the upper limit is reached, then step (S4) is executed.

[0077] (S4) Adjusting the drag optimization target and training set: If the electromagnetic performance optimization in step (S304) reaches the upper limit of the iteration count but still fails to meet the electromagnetic performance design target, the metal ground topology output in step (S104) can be considered to have poor electromagnetic performance. Continuing to optimize the electromagnetic performance of the antenna with this metal ground topology would be too costly. To reduce the computational burden and accelerate the overall optimization convergence, the training set for drag performance optimization is adjusted. The metal ground design parameters and their corresponding drag values ​​output in step (S104) are deleted, and step (S103) is executed to perform a new round of drag performance optimization, resulting in a metal ground with a new topology. The electromagnetic performance of the antenna composed of the new metal ground is then re-optimized.

[0078] Using the above algorithm, a low-drag dual-polarized crossed dipole antenna is designed, with the number of iterations for optimizing drag performance i... a The upper limit is 32, and the number of iterations i for electromagnetic performance optimization is... em The upper limit is 65; the optimization target for wind resistance performance is t. wl =0.22N, and t wl Adjusting within the range [0.22, 0.385]N, the new t wl The value is taken as the previous value multiplied by 1.1. This method avoids setting the initial wind resistance optimization target too strictly, which would make optimization convergence more difficult. Figure 7 The result is given after i a =After two iterations, the topology of the metallic ground plane satisfies the wind resistance optimization objective. The optimal design parameters of the metallic ground plane are [6,7,2,2,6,5,1,5,4,12,5,26.87]. The wind resistance of the antenna model composed of this metallic ground plane is 0.23N. Within the operating frequency band of 3.3-5.0GHz, the optimization objective for the electromagnetic performance of the antenna is as follows:

[0079] 1)|S 11 |<-10dB,|S 22 |<-10dB;

[0080] 2)|S 21 |<-20dB;

[0081] 3) Gain greater than 5.5 dBi.

[0082] Let the penalty values ​​ω1, ω2, ω3, and ω4 be equal to 50, 50, 50, and 1, respectively. After iem After 19 iterations, the antenna design parameters [3,3,4,1,1,2,14,4,3,4.63,0.99,28.16,6.61,9.21,20.16] that meet the electromagnetic performance specifications are obtained. The antenna's radiating layer structure is as follows: Figure 8 As shown, the full-wave simulation results of the S-parameters and gain are as follows: Figure 9 As shown in the figure. The simulation results show that the final output of the entire optimization algorithm is a low-drag dual-polarized cross dipole antenna that satisfies both wind resistance and electromagnetic performance indicators.

[0083] This invention also discloses a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded onto the processor, it implements the steps of the above-mentioned parameterized topology design and optimization method, or the sequential multiphysics machine learning-assisted low-drag dipole antenna design method.

[0084] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A sequential multiphysics machine learning-assisted low-drag antenna design method, characterized in that, Includes the following steps: Based on the parametric topology design and optimization method, the topology of the antenna metallic ground and the topology of the dipole are changed by mesh selection design parameters, material property design parameters and complex image generation design parameters, respectively, in order to optimize the antenna's wind resistance and electromagnetic performance. Among these measures, the structural parameters of the antenna are increased to optimize electromagnetic performance. The parametric topology design and optimization method includes: setting an initial mesh image and initial material properties of the mesh; setting mesh selection design parameters for selecting the location and number of mesh points, and material property design parameters for changing the material properties of the selected mesh points and their adjacent mesh points; setting complex image generation design parameters for generating more complex mesh images based on the mesh image with changed material properties, including parameters for determining the combination method of mesh image transformation, parameters for determining the combination order of the original mesh image and the transformed mesh image, and parameters for the number of combinations; for image rotation transformation, parameters for determining the rotation angle are also included; for the more complex mesh image generated based on the above design parameters, the contour lines between meshes with different material properties are extracted, and the topology structure corresponding to a set of design parameters is obtained by modeling based on the contour lines; for the defined antenna performance optimization target, a surrogate model is established using machine learning methods to assist performance optimization, and the topology design parameters of the structural layer that meet the design target are obtained; A sequential antenna drag performance optimization and electromagnetic performance optimization design process is constructed. Machine learning methods are used to establish a surrogate model in the antenna drag performance optimization process to accelerate the design process and obtain the topology design parameters of the metal ground that meet the drag design target. After assembling a complete cross-dipole antenna by combining a metal ground plane, a radiating layer, and a coaxial feed line that meet the wind resistance design target, machine learning methods are used to accelerate the electromagnetic performance optimization process of the antenna. This yields a dipole topology and antenna structural parameters that meet the requirements for bandwidth, isolation, and gain, ultimately resulting in a dual-polarized cross-dipole antenna that satisfies both low wind resistance design requirements and electromagnetic performance requirements.

2. The sequential multiphysics machine learning-assisted low-drag antenna design method according to claim 1, characterized in that, The proxy model is trained using machine learning methods, and the training dataset of the proxy model is obtained by sampling within the range of all design parameters and then performing simulation.

3. The sequential multiphysics machine learning-assisted low-drag antenna design method according to claim 1, characterized in that, When used for parameterized topology design and optimization of antenna metallic ground, an antenna wind resistance performance optimization target is defined. A surrogate model based on the wind resistance value obtained from the projected area of ​​the metallic ground is established using machine learning methods to assist in performance optimization and obtain metallic ground topology design parameters that meet the design target.

4. The sequential multiphysics machine learning-assisted low-drag antenna design method according to claim 1, characterized in that, When used for parameterized topology design and optimization of antenna dipoles, an antenna electromagnetic performance optimization objective is defined. A surrogate model is established using machine learning methods to obtain S-parameters and gain based on design parameters and frequency to assist in performance optimization, thereby obtaining dipole topology design parameters that meet the design objective.

5. The sequential multiphysics machine learning-assisted low-drag antenna design method according to claim 4, characterized in that, After optimizing the wind resistance performance of the antenna's metallic ground, the metallic ground is combined with the mesh support structure. When optimizing the overall electromagnetic performance of the antenna, the design parameters include: the number of grids per row / column of the mesh support structure; the size of the region in the radiating layer that is not topologically optimized; the size of the gap between the design area and the edge of the dipole arm; the diameter of the radiating layer; the size of the support structure around the coaxial feed line; the angle of the top of the dipole arm; and the distance between the radiating layer and the metallic ground.

6. The sequential multiphysics machine learning-assisted low-drag antenna design method according to claim 1, characterized in that: The antenna's wind resistance performance is optimized based on prior knowledge, taking into account the wind resistance calculation formula: ； Where F w Represents wind resistance, C f It is the drag coefficient, A ref It is the size of the projected area, q p It is dynamic pressure and positively correlated with the square of wind speed; by utilizing the correspondence between the design parameters of the metal ground and the size of the projected area of ​​the metal ground, a surrogate model between the projected area and the wind resistance value is established through machine learning algorithms to improve the accuracy of wind resistance prediction.

7. The sequential multiphysics machine learning-assisted low-drag antenna design method according to claim 1, characterized in that: The design parameters for electromagnetic performance optimization include both dipole topology design parameters and antenna structural parameters; the electromagnetic performance optimization of the antenna combines dipole topology optimization and antenna structural parameter optimization; during the electromagnetic performance optimization process, machine learning methods are used to establish a surrogate model to replace time-consuming full-wave simulation, thereby accelerating the antenna design process.

8. The sequential multiphysics machine learning-assisted low-drag antenna design method according to claim 1, characterized in that: In the sequential optimization design process, the output of antenna drag performance optimization is the input of electromagnetic performance optimization. Based on the drag performance optimization, the optimal topology design parameters of the metallic ground are obtained, and the metallic ground is modeled in the full-wave simulation software. This metallic ground, together with the radiating layer and coaxial feeder, forms the antenna model for the next stage of electromagnetic performance optimization. The output of antenna electromagnetic performance optimization affects whether the drag performance optimization process is restarted. In order to accelerate the optimization design of low drag antennas, antennas that still cannot meet the electromagnetic performance optimization target after reaching the upper limit of the optimization iteration number will be discarded, and drag performance optimization will be restarted to obtain a new metallic ground topology.

9. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is loaded into the processor, it implements the steps of the method according to any one of claims 1-8.

Images