Antenna artificial intelligence design system and method, electronic equipment and storage medium

Abstract

The invention discloses an antenna artificial intelligence design system and method, electronic equipment and a storage medium, and belongs to the technical field of antenna design. The system executes stored instructions through a processor, receives a design target, generates a geometric structure image, performs boundary correction based on an edge effect physical model to obtain an electrical size image, and generates a candidate antenna structure under the condition of the design target by using a conditional generative adversarial network. A multi-modal neural network is adopted to carry out fusion processing on multi-source features such as geometric structures, contours, electrical dimension images and key parameters, S parameters are predicted, and iterative optimization is carried out until design requirements are met. According to the method, a physical model is integrated to improve the prediction precision and generalization ability, joint automatic optimization of the antenna structure and size is realized, the design period is shortened to be within 3 hours from traditional 72 hours, and the problems that traditional design efficiency is low, and an existing AI method is insufficient in degree of freedom and ignores a physical mechanism are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of antenna design technology, and in particular to an artificial intelligence design solution for antennas that integrates physical models and multimodal neural networks, including systems, methods, electronic devices, and storage media. Background Technology

[0002] Planar microstrip antennas, as key components of modern wireless communication systems, play an irreplaceable role in 5G communication, the Internet of Things, and satellite navigation. With the continuous improvement of technical standards, more stringent requirements are being placed on their performance indicators and miniaturization designs, highlighting the limitations of traditional design methods. The current industry-wide design process relies heavily on engineers' personal experience and iterative iterations using full-wave electromagnetic simulation software (such as HFSS): designers first determine the initial structure and dimensional parameters of the antenna based on experience, then verify its performance using simulation software, and finally manually adjust the design based on the simulation results, repeating this cycle until the specifications are met. This highly manual design model not only makes the entire cycle take tens of hours or even days, but also makes the design quality closely related to the engineer's experience level, resulting in poor consistency and low reusability, severely hindering the rapid development and iteration of RF front-end products.

[0003] To improve the level of design automation, academia and industry have begun to explore the introduction of artificial intelligence technology into the field of antenna design. Some existing AI-assisted solutions attempt to use convolutional neural networks to model and predict specific parameters such as the resonant frequency of antennas, or combine neural networks with optimization algorithms such as simulated annealing for parameter tuning. However, these methods generally suffer from fundamental flaws: they can mostly only optimize a limited number of parameters under a preset, fixed antenna topology, and cannot achieve independent innovation and evolution of antenna geometry. In essence, they have not broken through the traditional design thinking framework, and the degree of design freedom remains severely limited.

[0004] A deeper problem lies in the fact that these AI models merely focus on learning the superficial data mapping relationship between design parameters and performance indicators, completely ignoring the underlying electromagnetic physics laws governing antenna operation. This purely data-driven modeling approach causes the model's prediction accuracy to drop sharply when faced with design scenarios not covered by the training data, resulting in a significant lack of generalization ability.

[0005] Furthermore, most existing solutions can only handle a single type of design data, such as structural images or parameter vectors, failing to fully explore and utilize the inherent correlations and complementarities between different types of data (such as geometry, electromagnetic properties, and dimensional parameters), further limiting their performance ceiling and application scope.

[0006] In summary, there is an urgent need in this field for an innovative solution that can overcome the aforementioned limitations. This solution should be able to achieve collaborative and autonomous optimization of antenna geometry and dimensional parameters, deeply integrate electromagnetic physics mechanisms to improve the accuracy and generalization ability of the model, and fully utilize the intrinsic value of multi-source heterogeneous data, thereby truly realizing efficient, accurate, and automated artificial intelligence design of antennas. Summary of the Invention

[0007] The technical problem to be solved by this invention is to overcome the problems of low efficiency, insufficient design freedom, and limited prediction accuracy and generalization ability of existing antenna design methods.

[0008] To solve the above problems, the present invention adopts the following technical solution: In a first aspect, the present invention provides an antenna artificial intelligence design system, comprising a processor and a memory, wherein the memory stores program instructions that run on the processor, and the processor, when executing the instructions, implements: The input unit is used to receive antenna design targets containing target performance indicators; The data processing unit is used to generate a binary geometric structure image of the antenna based on the key size parameters, extract the edges of the geometric structure image to obtain a contour structure image, and perform boundary correction on the geometric structure image based on the extension calculated by the edge effect physical model to generate an electrical size image. The generation unit includes a conditional generative adversarial network, whose generator is used to receive a conditional embedding consisting of a potential vector and the design objective, and output a candidate antenna geometry image. The performance prediction unit includes a multimodal neural network, which has three image branches that process the geometric structure image, the contour structure image, and the electrical dimension image respectively, and a parameter branch that processes the key dimension parameters. The features output by each branch are spliced ​​and fused to output a prediction vector characterizing the antenna S-parameters. An optimization unit is used to compare the prediction vector output by the performance prediction unit with the target performance index. If the target performance index is not met, the generation unit is triggered to generate a new candidate antenna geometry image for iterative optimization.

[0009] As an optional implementation, the data processing unit performs edge extraction on the geometric structure image using the Sobel operator.

[0010] As an optional implementation, the discriminator of the conditional generative adversarial network is used to simultaneously receive candidate images or real images and the design target, and perform a real vs. fake determination.

[0011] As an optional implementation, each of the three image branches of the multimodal neural network includes a convolutional layer and a pooling layer, and the parameter branch is composed of a fully connected layer.

[0012] As an optional implementation, the multimodal neural network uses mean absolute error as the loss function.

[0013] Secondly, the present invention provides an antenna artificial intelligence design method, employing the antenna artificial intelligence design system described in the first aspect above, comprising the following steps: Step 1. Receive the antenna design target, which includes the target performance indicators; Step 2. Generate a binary geometric structure image of the antenna, extract its contour structure image, and perform boundary correction on the geometric structure image based on the edge effect physical model to generate an electrical dimension image; Step 3. Using the design objective as a condition, generate candidate antenna geometry images using a conditional generative adversarial network generator; Step 4. Extract the features of the candidate geometric structure image and its corresponding contour structure image, electrical dimension image and key dimension parameters respectively through a multimodal neural network, and stitch and fuse the obtained intermediate features to predict the S-parameters of the antenna. Step 5. Compare the predicted S-parameters with the target performance indicators. If the target is not met, repeat steps 3 to 5 until the design target is met.

[0014] As an optional implementation, in step 2, the boundary correction is achieved by calculating the equivalent dielectric constant and the radiation electric dimension extension length using the edge effect calculation formula.

[0015] As an optional implementation, in step 4, the training of the multimodal neural network uses a cosine annealing strategy to adjust the learning rate.

[0016] Thirdly, the present invention provides an electronic device, comprising: a processor and a memory; The memory is used to store computer programs; The processor is used to execute the computer program to implement the antenna artificial intelligence design method described in the second aspect.

[0017] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the antenna artificial intelligence design method as described in the second aspect.

[0018] This invention achieves joint automated optimization of antenna geometry and dimensional parameters by constructing an intelligent design framework that coordinates conditional generative adversarial networks and multimodal neural networks. This breaks the limitations of traditional preset topologies, significantly expands design freedom, and makes it possible to generate novel, irregular antenna slotted structures. Furthermore, by incorporating physical models such as edge effects into the data preprocessing stage, an electrical dimension image that accurately characterizes electromagnetic space properties is generated. This effectively compensates for the inherent differences between geometric and electromagnetic spaces. This innovative approach allows the model to learn essential characteristics that better conform to physical laws, resulting in a significant performance improvement of approximately 8% reduction in test loss and a 41% increase in convergence speed, greatly enhancing the model's prediction accuracy and generalization ability.

[0019] Furthermore, by constructing a complete closed-loop feedback process of "generation-prediction-optimization", the traditional design mode that relies on repeated manual adjustments is transformed into a fully automated intelligent iteration, achieving a qualitative leap in design efficiency. The typical design cycle is shortened from more than 72 hours in the traditional method to less than 3 hours, with an efficiency improvement of more than 80%, which greatly reduces R&D costs and time.

[0020] Ultimately, the system utilizes a multimodal neural network to process multi-source information such as geometric structures, contours, electrical dimensions, and key parameters in parallel, and performs effective feature-level fusion to fully exploit the complementary value between different modal data. This allows it to stably achieve the performance requirement of S11 ≤ -10dB in multiple target frequency bands, including 2.45GHz, with an error of less than 3% between the prediction results and high-precision electromagnetic simulation results, demonstrating excellent engineering application potential and broad scenario adaptability. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0022] Figure 1 This is a schematic diagram of the overall architecture of the antenna artificial intelligence design system provided in an embodiment of the present invention; Figure 2 This is a complete flowchart of the antenna artificial intelligence design method provided in the embodiments of the present invention; Figure 3 This is a schematic diagram illustrating the process of constructing a multimodal training dataset in an embodiment of the present invention; Figure 4 This is a specific structural diagram of the multimodal neural network in an embodiment of the present invention; Figure 5 This is a training structure diagram of the conditional generative adversarial network in an embodiment of the present invention; Figure 6This is a graph showing the change in the loss function of the multimodal neural network during the training process in an embodiment of the present invention; Figure 7 This is a comparison diagram of the positive prediction values ​​and true values ​​of the antenna S-parameters by the multimodal neural network in an embodiment of the present invention; Figure 8 This is a performance comparison chart of the antenna layout designed by this system and its predicted results with the simulation results in an embodiment of the present invention. Detailed Implementation

[0023] The following examples are used to illustrate the present invention, but are not intended to limit the scope of the invention.

[0024] Example 1: Please see Figure 1 This embodiment provides an artificial intelligence antenna design system. This system integrates an electromagnetic physics model with a multimodal deep learning network to achieve automated and intelligent design of planar microstrip antennas. Specifically, the system includes an input unit, a data processing unit, a generation unit, a performance prediction unit, and an optimization unit. These functional units are implemented collaboratively by a processor and a memory. The processor, as the core of the system's computation, can be a central processing unit, a graphics processing unit, or a tensor processor, or other hardware with parallel computing capabilities. The memory includes random access memory and read-only memory, used for temporarily storing computational data and permanently storing program instructions. When the system starts, the processor reads and executes preset program instructions from the memory, sequentially calling each functional unit to complete the antenna design task.

[0025] The input unit is responsible for receiving the antenna design goals set by the user. These goals include basic design requirements such as operating frequency band and performance indicators. For example, in the case of WiFi antenna design, the design goals typically include an operating frequency of 2.45 GHz and a performance requirement of S11 ≤ -10 dB.

[0026] After receiving the design parameters from the input unit, the data processing unit first generates the corresponding binary geometric structure image based on the key dimension parameters. (See also...) Figure 3 The diagram illustrates the dataset construction process. The geometric structure image is represented using black-and-white binary representation, where white pixels represent the metallic radiating portion of the antenna, and black pixels represent the dielectric substrate. This image representation method clearly shows the slotted structure and metallic distribution characteristics of the antenna, and it is stored in an image convolution-compatible format, facilitating subsequent neural network processing. Next, the data processing unit extracts edge features from the geometric structure image. This process is implemented using the Sobel operator, which can effectively detect gradient changes in the image, thereby extracting the antenna's contour information and forming a contour structure image.

[0027] To incorporate electromagnetic physics properties into the design process, the data processing unit also generates an electrical dimension image based on a physical model of edge effects. This step is achieved by calculating the equivalent dielectric constant and the radiation electrical dimension extension length, where: Equivalent dielectric constant The calculation formula is: ; Radiated electrical dimension extension length The calculation formula is: .

[0028] In the above formulas, Represents the relative permittivity of the dielectric substrate. For the thickness of the dielectric substrate, This refers to the microstrip transmission line width. The extension calculated using the above formula is used to correct the boundaries of the geometric image, resulting in an electrical dimension image that more accurately characterizes the antenna's electromagnetic radiation characteristics. In practical implementations, the dielectric substrate is typically made of FR4 material, which has a relative permittivity of... It is 4.4, and the thickness is... These parameters, fixed at 1.6 mm, provide the necessary physical basis for electrical dimension calculations.

[0029] Please see Figure 5 The generation unit contains a conditional generative adversarial network (GAN), which consists of a generator and a discriminator. The generator receives conditional inputs based on the design objectives and fuses these conditions with random latent vectors using a transposed convolutional network structure, progressively upsampling to generate candidate antenna geometric images. The discriminator employs a convolutional network structure, simultaneously receiving real antenna images or candidate images generated by the generator along with the corresponding design objective conditions, and performs training to distinguish between real and fake images. This GAN architecture ensures that the generated antenna structures both conform to the design objectives and maintain structural rationality and diversity. Figure 5 It also demonstrates the adversarial iterative mechanism between the generator and discriminator during the training process. By optimizing the binary cross-entropy loss function, the generator is able to gradually generate antenna structure images that conform to the real distribution.

[0030] Please see Figure 4The performance prediction unit employs a multimodal neural network architecture, comprising three image processing branches and one parameter processing branch. The first image processing branch specifically handles contour structure images, extracting spatial features through convolutional layers Conv_C1 and Conv_C2, followed by feature compression via a fully connected layer FC_C. The second image processing branch processes geometric structure images, extracting contour features through convolutional layers Conv_I1 and Conv_I2, then through a max-pooling layer MaxPool_I, and finally outputting the feature vector via a fully connected layer FC_I. The third image processing branch handles electrical dimension images, progressively extracting deep features through three convolutional layers Conv_E1, Conv_E2, and Conv_E3, followed by feature encoding via a max-pooling layer MaxPool_E and a fully connected layer FC_E. These image processing branches also include other layers, such as activation layers or normalization layers.

[0031] The parameter processing branch specifically handles key size parameters, learning the mapping relationship between parameters and performance through fully connected layers FC_P1, FC_P2, and FC_P3. The features extracted from each branch are concatenated in the fusion layer to form a unified feature representation, and finally, a predicted vector representing the antenna S-parameters is output through the fully connected layer.

[0032] The optimization unit is responsible for coordinating the iterative optimization of the entire design process. This unit compares the S-parameter prediction vector output by the performance prediction unit with the target performance indicators set by the input unit in real time. When the prediction result does not meet the design requirements, it automatically triggers the generation unit to generate new candidate antenna geometry images, forming a closed-loop optimization mechanism. This iterative optimization process continues until an antenna scheme that meets all design objectives is obtained. Finally, the output unit outputs the optimal antenna structure image and the corresponding performance prediction results.

[0033] During the system training phase, sufficient data samples are needed to train the conditional generative adversarial network (GAN) and the multimodal neural network (MNN). This embodiment uses 15,000 sets of antenna sample data, divided into training and validation sets in an 8:2 ratio. The MNN is trained using the Adam optimizer with a learning rate of 0.0002 and 150 iterations. The GAN is trained using an image dataset with a batch size of 128 and momentum parameters β1=0.5 and β2=0.999. The network is planned to train for 1000 epochs, saving the generated sample visualization results every 400 batches and saving the model weights every 500 epochs. During training, the generator receives 50-dimensional truncated normal distribution noise as input and gradually upsamples it through transposed convolutional layers to generate 200×200 pixel grayscale images. The discriminator distinguishes between real and generated samples through a convolutional downsampling structure. Both continuously optimize through adversarial learning, gradually approximating the data distribution of the real antenna samples.

[0034] Please see Figure 6 Multimodal neural networks continuously adjust their internal parameters by optimizing a loss function, where the loss function used is the mean absolute error, and the specific calculation formula is as follows: ,in This represents the value of the mean absolute error loss function. Represents the first in the dataset The true S-parameter values ​​of each sample This represents the multimodal neural network for the first... The predicted S-parameter values ​​output for each sample, and This represents the total number of samples contained in a training batch. The smaller the value, the higher the prediction accuracy of the network. Figure 6 The study demonstrates that the loss function steadily decreases with increasing iterations during training, indicating that the multimodal neural network has good convergence performance.

[0035] In summary, the antenna AI design system described in this embodiment achieves full automation of the antenna design process through the synergistic effect of the aforementioned units. The system first transforms the user's design goals into specific structural generation conditions, then generates diverse candidate structures through a conditional generative adversarial network (GAN), rapidly evaluates the performance of these structures using a multimodal neural network, and finally selects the optimal design scheme through a closed-loop optimization mechanism. This design approach, based on the fusion of physical models and deep learning, not only significantly improves design efficiency, shortening the traditional design cycle of tens of hours to less than 3 hours, but also enhances the model's predictive accuracy and generalization ability by introducing electromagnetic physical characteristics, providing strong technical support for the antenna design of modern wireless communication devices.

[0036] Example 2: Please see Figure 2 This embodiment provides an artificial intelligence design method for antennas. This method can be implemented using the system described in Embodiment 1, or it can be executed on other computing devices. The method includes the following steps: First, step 1 is executed, receiving the antenna design goals input by the user. These goals include specific indicators such as operating frequency band and S-parameter performance requirements. In practical applications, design goals can be set for different communication standards. For example, for Bluetooth antenna design, the operating frequency can be set to 2.4GHz; for 5G terminal antennas, the operating frequency may cover multiple frequency bands such as 3.5GHz and 4.9GHz. Performance indicators typically include parameters such as return loss, radiation efficiency, and gain. Among these, the S11 parameter is the most basic and important performance indicator, generally required to be ≤-10dB within the design frequency band.

[0037] Step 2 generates the initial geometry of the antenna and processes multimode data. Based on the input key size parameters, a binary geometric structure image is first generated. This image is stored in matrix form, and the image size can be flexibly set according to accuracy requirements. For example, a resolution of 256 pixels × 256 pixels can be optionally used to ensure sufficient detail representation. Next, edge extraction is performed on the geometric structure image. This process is implemented using the Sobel operator, which contains convolution kernels in both horizontal and vertical directions, effectively detecting edge features in the image and forming a contour structure image. Then, an electrical dimension image is generated based on the edge effect physical model. Specifically, the boundaries of the geometric structure image are corrected by calculating the equivalent dielectric constant and the radiated electrical dimension extension length. The calculation of the extension length involves the electromagnetic property parameters of the dielectric substrate. For example, when using FR4 material, the relative dielectric constant... The value is 4.4, and the substrate thickness is... It is 1.6mm.

[0038] Step 3 uses a conditional generative adversarial network (GAN) to generate candidate antenna structures. The structure of this GAN is as follows: Figure 5 As shown, the generator receives conditional inputs consisting of design targets and outputs candidate antenna geometry images. The generator employs a transposed convolutional network structure, which processes and transforms the conditional information to progressively generate candidate images. Correspondingly, the discriminator uses a convolutional network structure, simultaneously receiving candidate images or real images and the design targets, and performing authenticity checks. This mechanism ensures a high degree of matching between the generated structures and the design targets. During training, the generator and discriminator perform adversarial alternating optimization based on a binary cross-entropy loss function. Through multiple iterations, the generator eventually produces images that conform to the actual antenna structure distribution and meet the design conditions.

[0039] Step 4 predicts antenna performance using a multimodal neural network. The candidate geometric structure image generated in Step 3, along with its corresponding contour structure image, electrical dimension image, and key dimensional parameters, is input into the multimodal neural network. The network's four branches process different types of input data in parallel: the geometric structure image branch extracts metal distribution features through convolutional layers; the contour structure image branch focuses on edge and shape features; the electrical dimension image branch learns electromagnetic field distribution characteristics; and the parameter branch directly processes numerical dimensional parameters. The intermediate features extracted by each branch are then concatenated in a pre-defined fusion layer. Figure 7 The comparison chart of the predicted and actual values ​​of the multimodal neural network shown demonstrates the high consistency between the network prediction results and the HFSS simulation results, proving the accuracy of the method.

[0040] Step 5 involves performance evaluation and iterative optimization. The S-parameter curves predicted by the multimodal neural network are compared with the design target, and the performance difference index is calculated. If the prediction results do not meet the design requirements, for example, S11 > -10dB at the target frequency of 2.45GHz, an optimization signal is generated based on the prediction error, the input conditions of the conditional generative adversarial network are adjusted, and steps 3 to 5 are repeated.

[0041] Various strategies can be employed during the optimization process, such as randomly perturbing the latent vector, adjusting the weight coefficients of the conditional inputs, or introducing diversity promotion mechanisms, to ensure that the globally optimal solution is found. When the predicted performance of the candidate antenna structure meets all design objectives, the final antenna design scheme is output, including the geometric structure image, key dimensional parameters, and predicted S-parameter curves.

[0042] In the specific implementation of the method, training the multimodal neural network is a crucial step. Training data can be generated using HFSS electromagnetic simulation software. The simulation settings cover a frequency scanning range of 1 GHz to 6 GHz, with a step size of 0.01 GHz to ensure sufficiently detailed S-parameter curves. The training process uses the mean absolute error as the loss function, selects Adam as the optimizer, and adjusts the learning rate using a cosine annealing strategy. The initial learning rate is set to 0.01, and the minimum learning rate is 0.000001. The training cycle is set to 500 epochs, and the batch size is determined based on the GPU memory capacity, for example, optionally set to 32 or 64.

[0043] Furthermore, to prevent overfitting, an early stopping mechanism can be introduced during training, which automatically terminates training when the validation set loss no longer decreases for several consecutive epochs.

[0044] The antenna AI design method described in this embodiment establishes a complete automated antenna design process through innovative approaches such as multimodal data fusion and physical model embedding. This method first transforms the design objective into specific generation conditions, then uses a conditional generative adversarial network to generate diverse antenna structures, followed by rapid performance evaluation using a multimodal neural network, and finally selects the optimal solution through a closed-loop optimization mechanism. This AI-based design method not only significantly improves design efficiency but also enhances the reliability and practicality of the model by incorporating electromagnetic physics knowledge, providing a new technical path for the field of antenna design. Figure 8 The comparison chart of antenna design results shown clearly demonstrates the high degree of agreement between the antenna layout designed by this method and its predicted performance and the simulation results, verifying the effectiveness and accuracy of the method.

[0045] Example 3: This embodiment provides an electronic device that integrates the antenna artificial intelligence design system described in Embodiment 1, and can be used to implement the antenna artificial intelligence design method described in Embodiment 2. The core components of the electronic device include a processor, memory, input devices, and output devices, which transmit data and instructions via a system bus.

[0046] The processor, as the computing center of the device, can be a multi-core central processing unit, a graphics processing unit, or a tensor processing unit specifically designed for neural network computing. It is responsible for performing various computational tasks in the antenna design process, including image processing, neural network inference, and optimization algorithms.

[0047] Memory includes volatile memory and non-volatile memory. Volatile memory, such as dynamic random access memory, is used for temporary storage of intermediate data during the computing process, while non-volatile memory, such as solid-state drives or flash memory, is used for permanent storage of operating systems, applications, and antenna design databases.

[0048] Input devices provide users with interfaces to interact with the system, including but not limited to keyboards, mice, touchscreens, and voice input devices. Designers can use these devices to input antenna design goals, adjust parameter settings, and monitor design progress. Output devices are used to display design results and system status, including high-resolution displays, printers, and audio output devices. The displays are used to visualize generated antenna structure images, predicted S-parameter curves, and various graphs of the optimization process. In specific implementations, electronic devices may also include network interfaces for connecting to remote servers and databases, enabling cloud synchronization of design data and collaborative design capabilities.

[0049] The software architecture of an electronic device comprises an operating system layer, an application programming interface (API) layer, and an antenna design application layer. The operating system layer provides basic hardware management, memory allocation, and task scheduling functions, and can be a mainstream operating system such as Windows, Linux, or macOS. The API layer contains various mathematical computation libraries and deep learning frameworks, such as NumPy, TensorFlow, or PyTorch, providing efficient numerical computation and neural network operation support for antenna design algorithms. The antenna design application layer implements the various functional units described in Example 1, including input units, data processing units, generation units, performance prediction units, and optimization units. These units are organized as software modules and communicate and exchange data through clearly defined interfaces.

[0050] In the actual deployment of electronic devices, when the processor executes the program instructions stored in the memory, it first initializes the antenna design environment and loads the pre-trained neural network model and physical model parameters. Upon receiving a design task submitted by the user through an input device, the processor executes the design process sequentially according to the steps described in Example 2, outputting intermediate results to the display for user monitoring at key nodes. For computationally intensive tasks, such as neural network inference and electromagnetic simulation, the processor utilizes hardware accelerators such as GPUs or TPUs to improve computational efficiency. After the design is completed, the processor saves the final antenna design scheme to the memory and simultaneously displays a detailed design report to the user through an output device, including antenna structure diagrams, dimensional parameters, performance prediction curves, and a comparative analysis with traditional design methods.

[0051] The electronic device described in this embodiment provides antenna designers with an efficient and convenient intelligent design tool through the coordinated optimization of hardware and software. The device integrates advanced deep learning algorithms and electromagnetic physics models, transforming the traditionally knowledge- and experience-dependent antenna design process into an automated computational task, significantly lowering the design threshold and improving design quality and efficiency. Simultaneously, the device's modular architecture ensures system flexibility and scalability, allowing for the customization of functional modules according to different application requirements. It adapts to antenna design tasks across different frequency bands, from microwave to millimeter waves, demonstrating broad application prospects and practical value.

[0052] Example 4: This embodiment provides a computer-readable storage medium storing a computer program. When the program is executed by a processor, it implements the antenna artificial intelligence design method described in Embodiment 2. The physical form of the storage medium can be a solid-state drive, USB flash drive, optical disc, SD card, or server hard drive array, etc., and its storage capacity is determined according to actual needs. For example, a 512GB solid-state drive can be optionally used to store program code, training data, and design results. The file system format of the storage medium is selected according to the platform used, such as NTFS format in Windows systems and EXT4 format in Linux systems, to ensure reliable data storage and efficient read / write.

[0053] The computer program stored in the storage medium contains multiple functional modules organized according to the antenna design flow, including a data input module, an image generation module, a physical calculation module, a neural network inference module, and an optimization output module. The data input module is responsible for parsing the user-input design goals, verifying the validity of parameters, and converting them into the system's internal representation format. The image generation module constructs the antenna's geometric structure image based on key dimensional parameters, supporting various common antenna topologies such as rectangular patches, circular patches, E-shaped and U-shaped slotted structures. The physical calculation module implements the core algorithms of the edge effect physical model, including functions such as equivalent dielectric constant calculation, radiated electrical dimension extension length calculation, and boundary pixel correction, ensuring that the generated electromagnetic dimension image accurately reflects the antenna's radiation characteristics.

[0054] The neural network inference module integrates pre-trained conditional generative adversarial network (GAN) models and multimodal neural network (NN) models, which are stored as files on the medium. The GAN model files contain the network structure definitions and weight parameters for the generator and discriminator, while the NN model files store detailed configurations and trained parameters for the four processing branches. During program execution, these model files are loaded into memory and inference computation is performed using the corresponding deep learning frameworks. The optimization output module is responsible for coordinating the iterative optimization of the design process, monitoring the gap between performance predictions and design goals, and dynamically adjusting the input conditions of the generative network until a design solution that meets the requirements is obtained.

[0055] The storage medium also includes auxiliary data files, such as an antenna design specification library, a material parameter library, and a typical design case library. The design specification library stores antenna performance requirements for different communication standards, such as frequency band allocations and S-parameters for standards like WiFi, Bluetooth, and 5G. The material parameter library includes the electromagnetic properties of various commonly used dielectric substrates, including parameters such as the relative permittivity, loss tangent, and temperature coefficient of materials like FR4, Rogers series, and PTFE. The typical design case library provides verified successful design examples, offering users references and design starting points.

[0056] To ensure portability and compatibility, the software in the storage medium is implemented using a cross-platform development framework. The core algorithms can be written in C++ and Python, and the user interface adapts to the operating environment, providing rich interactive functionality on desktop computers and a simple operating interface on embedded devices. The program installation package includes a complete list of dependency libraries and automatic configuration scripts, supporting one-click deployment and rapid upgrades. Furthermore, the storage medium provides detailed API documentation and development guidelines for easy integration with other systems and functional expansion.

[0057] The computer-readable storage medium described in this embodiment transforms advanced antenna AI design methods into executable software, enabling antenna design engineers to easily access AI-assisted design capabilities. The complete program code and pre-trained models stored in the medium eliminate the technical barrier for users to develop their own algorithms, while the rich design library and material database provide comprehensive support for various application scenarios. This hardware-software separation implementation not only reduces usage costs but also facilitates system maintenance and updates. Design performance can be improved or new functions added by replacing or updating the program files in the storage medium, providing a practical technical foundation for the intelligent transformation of the antenna design industry.

Claims

1. An antenna artificial intelligence design system, characterized by, The system comprises a processor and a memory, the memory stores program instructions for running on the processor, and the processor implements the following steps when executing the instructions: An input unit is configured to receive an antenna design target comprising a target performance indicator; A data processing unit is configured to generate a binary geometry image of an antenna based on the critical dimension parameter, perform edge extraction on the geometry image to obtain a contour structure image, and perform boundary correction on the geometry image based on an extension calculated by an edge effect physical model to generate an electrical size image; A generation unit comprises a conditional generative adversarial network, a generator of which is configured to receive a conditional embedding composed of a latent vector and the design target, and output a candidate antenna geometry image; A performance prediction unit comprises a multi-modal neural network having three image branches respectively processing the geometry image, the contour structure image and the electrical size image, and a parameter branch processing the critical dimension parameter, and the features output by each branch are spliced and fused to output a prediction vector representing the S parameter of the antenna; An optimization unit is configured to compare the prediction vector output by the performance prediction unit with the target performance indicator, and if the requirement is not met, trigger the generation unit to generate a new candidate antenna geometry image for iterative optimization.

2. The antenna artificial intelligence design system of claim 1, wherein: The edge extraction on the geometry image by the data processing unit is implemented by a Sobel operator.

3. The antenna artificial intelligence design system of claim 1, wherein: A discriminator of the conditional generative adversarial network is configured to receive a candidate image or a real image and the design target simultaneously, and perform authenticity discrimination.

4. The antenna artificial intelligence design system of claim 1, wherein: The three image branches of the multi-modal neural network each comprise a convolution layer and a pooling layer, and the parameter branch is composed of a fully connected layer.

5. The antenna artificial intelligence design system of claim 1, wherein: The multi-modal neural network adopts mean absolute error as a loss function.

6. The antenna artificial intelligence design method of any one of claims 1 to 5, characterized by, The method comprises the following steps: Step 1. Receiving an antenna design target comprising a target performance indicator; Step 2. Generating a binary geometry image of an antenna, extracting a contour structure image thereof, and performing boundary correction on the geometry image based on an edge effect physical model to generate an electrical size image; Step 3. Generating a candidate antenna geometry image through a generator of a conditional generative adversarial network with the design target as a condition; Step 4. Extracting features of the candidate geometry image, its corresponding contour structure image, electrical size image and critical dimension parameter through a multi-modal neural network, splicing and fusing the obtained intermediate features, and then predicting the S parameter of the antenna; Step 5. Comparing the predicted S parameter with the target performance indicator, and if the target is not met, repeating steps 3 to 5 until the design target is met.

7. The antenna artificial intelligence design method of claim 6, wherein: In step 2, the boundary correction is achieved by calculating the equivalent permittivity and the extension length of the radiation electrical size through an edge effect calculation formula. 8.The antenna artificial intelligence design method of claim 6, wherein, In step 4, the training of the multi-modal neural network adopts a cosine annealing strategy to adjust the learning rate.

9. An electronic device, characterized by The system comprises: a processor and a memory; the memory is configured to store a computer program; the processor is configured to execute the computer program to implement the antenna artificial intelligence design method according to any one of claims 6 to 8.

10. A computer readable storage medium, characterized in that, The storage medium has stored thereon a computer program which, when executed by a processor, implements the antenna artificial intelligence design method according to any one of claims 6 to 8.

Images