Future Technologies
AI-Driven Innovations in RF and Antenna Design
This paper widely explores the application of AI technologies in crucial phases of RF and antenna design.

By Huawei Wireless Technology Lab: Guangjian Wang, Lingjun Yang, Jimmy Jian, Chandan Roy, Li Pan, Guolong Huang, Hua Cai, Wen Tong
1 Introduction
In the information age, radio frequency (RF) and antenna technologies are the cornerstone of modern wireless communications systems, connecting people and the world. Without these technologies, we would be unable to communicate through mobile devices or keep in touch with our family and friends. Satellite communications would be paralyzed, making global information transmission impossible. The Internet of Things (IoT) would become an unrealizable fantasy, and smart devices would never be able to communicate with each other. RF and antenna design is a crucial component of a wireless communication system because the performance of RF devices and antennas has a significant impact on the quality and efficiency of the entire system.
However, the rapid development of technology and the growing demand for communications have created several challenges for RF and antenna technologies. These challenges include improving the transmission efficiency of RF signals, designing more compact and efficient antennas, and optimizing the use of increasingly scarce spectrum resources. Conventional design approaches rely on experience and trial-and-error tests to address these challenges. These approaches are laborious and computationally inefficient, especially when there are a large number of co-dependent component parameters. Rapidly evolving artificial intelligence (AI) technologies have been widely applied in RF and antenna design due to their advanced data processing and learning capabilities, introducing innovative ideas and approaches to RF and antenna design.
In this paper, we explore innovative techniques to improve the performance of RF and antenna technologies. We conducted in-depth research on key fields such as RF circuit design, antenna design, and electromagnetic simulation to unlock the potential of AI in these fields. Through experiments and analysis, we investigated the use of AI technologies in simulating complex electromagnetic phenomena and achieved automatic adjustment of simulation parameters through reinforcement learning, significantly improving simulation efficiency. Figure 1 illustrates the composition of this paper. Section 2 outlines different AI technologies and AI models that can be used in RF and antenna design. Section 3 exemplifies the use of AI in RF circuit design, antenna design, and electromagnetic simulation. Section 4 discusses the unique advantages of AI in RF circuit and antenna design. Section 5 provides implementation examples of AI-based circuit design. Section 6 summarizes the challenges and prospects of AI in RF circuit design. Section 7 offers a conclusion to our work.

Figure 1 Composition of the paper
2 AI Models for RF and Antenna Design
AI is redefining RF and antenna design. AI technologies, especially machine learning (ML) and deep learning (DL), enable machines to learn and solve complex problems by imitating human cognitive processes and identifying data patterns. These technologies have been applied in RF and antenna design to improve design efficiency and enhance the design results, especially in solving high-dimensional, nonlinear, and multi-physical field coupling problems.
Figure 1 Composition of the paper
In RF design, feedforward neural networks (FFNNs), such as multilayer perceptrons (MLPs) and radial basis function (RBF) networks, are excellent solutions to non-dynamic modeling problems. Wavelet neural networks (WNNs) can solve problems with high nonlinearity or sharp changes due to their local property of hidden neurons, facilitating model training and improving model precision. Extreme learning machines (ELMs) are FFNNs with only one hidden layer that can learn quickly and yield good performance in complex electromagnetic parameter modeling scenarios (especially when only small-sized training datasets are available). An ELM can accurately simulate the behavior of RF circuits to optimize the design of filters and amplifiers. Dynamic neural networks, recurrent neural networks (RNNs), and time-delay neural networks (TDNNs) play a significant role in characterizing the dynamic behavior of a non-linear device or circuit in the time domain [6]. Deep neural networks (DNNs) provide an advanced solution to complex modeling problems due to their multi-layer structure [7]. Generative adversarial networks (GANs) demonstrate significant advantages in the generation of novel designs [8, 9]. Knowledge-based neural networks (KBNNs) [10] use the current equivalent circuit and empirical model in computer-aided design (CAD) for microwave components, eliminating the need to prepare a large amount of training data and improving the extrapolation capability of the model. These technologies can be integrated to accelerate the design process and advance the overall design performance and reliability.
3 Application of AI in RF and Antenna Design
3.1 RF Circuit Design
AI is transforming RF circuit and antenna design by accelerating the design process and significantly improving circuit performance. For instance, in filter and coupler design, AI can be used to accurately calculate and optimize frequency responses, coupling attenuation characteristics, and group delays, meeting the diverse range of frequency and power selection requirements In amplifier design, AI technologies introduce intelligent algorithms to optimize key parameters such as the gain and noise coefficient, achieving excellent amplifier performance under changing working conditions [15]. In oscillator design, AI can be used to design highly stable oscillators with low phase noise, delivering a high spectral purity even in complex electromagnetic environments. AI can also be employed to accurately calculate the impedance matching network, significantly improving the transmission efficiency and overall performance of RF circuits.
In filter design, numerous techniques can meet the performance requirements of filters. These techniques can be selected based on the application scenarios. For instance, microstrip and stripline technologies can be optimally integrated with planar circuits. Conventional waveguide technology is suitable for high-frequency applications with a low transmission loss. The innovative substrate-integrated waveguide (SIW) can be used in high-frequency applications due to its low transmission loss and the ability to be optimally integrated with planar system structures. The design parameters involved in these techniques vary depending on the filter structure. For instance, the length and width of a microstrip are major geometric variables of a microstrip filter. The diameter of a via and the distance between a via and the next via are major variables of an SIW filter. The aperture length and window length are key geometric variables of a waveguide filter. The impact of geometric variables on filter performance is often characterized or evaluated through ML technology. A well-trained ML model is expected to predict the impact of geometric variables on the filter response. Such a model can be used to quickly improve the filter design without a computationally expensive and inefficient electromagnetic model.
3.2 Antenna Design

Figure 2 Antenna design process
AI technologies enhance the efficiency and accuracy of antenna design. On the one hand, AI can be used to quickly generate antenna simulation results, eliminating the use of any model. On the other hand, AI can be used to fine-tune the size, shape, and structure of antennas or add specific structural elements, improving antenna performance in terms of gain, bandwidth, directivity, and other indicators. Additionally, AI can analyze the impacts of different polarization modes on signal transmission and reception, enabling engineers to select an optimal polarization mode, such as linear polarization, circular polarization, or elliptical polarization, based on scenario-specific requirements. This further improves signal transmission efficiency and reception quality. In antenna array design, AI algorithms can be used to design an optimal antenna layout, determine the number of elements, and optimize beam forming and control performance, significantly enhancing antenna directivity and gains. In multi-band antenna design, AI technologies take antenna parameters and performance requirements for different frequency bands into account, achieving advanced compatibility and efficient operation of antennas on multiple frequency bands — the diversity of spectrum resources is essential in modern communications systems. In conclusion, the integration and application of AI technologies have resulted in many innovative antenna design solutions. These solutions significantly improve design efficiency and system performance, laying a robust foundation for the evolution of wireless communications.
Figure 2 illustrates the process of AI-based antenna design. First, engineers select a basic geometric shape based on their experience to meet the performance requirements and then find the optimal design parameters. During parameter optimization, updating model parameters through full-wave electromagnetic simulation is the most time-consuming step. This step can be replaced by an ML-based surrogate model to save computing resources and accelerate the antenna design process. A surrogate model can be trained on the dataset obtained through full-wave simulation, a process that generates output parameters such as S and gain parameters on the operating frequency bands by changing the geometric shapes of antennas. In the training process, the result generated by the surrogate model gradually approaches the full-wave simulation result, eliminating the need for full-wave simulation. Most surrogate model-assisted antenna optimization approaches use the Gaussian Process (GP) surrogate model due to its easy implementation, processable analytical results, and ability to quantify uncertainties. Training of a surrogate model requires approximately 80% of datasets, while the remaining 20% is used to test model accuracy. If the test error does not meet the requirements, more data can be generated for training, or the surrogate model can be improved to repeat the GP process. After optimal device parameters are obtained through the surrogate model, the device needs to be verified and fine-tuned based on full-wave simulation to achieve optimal model accuracy and design performance. In antenna analysis, ML can be performed to predict various characteristics, such as radiation patterns and resonance frequencies, from simulation or measurement data. A complex surrogate model with multiple outputs can be integrated with multi-objective evolution algorithms to find the optimal antenna design that meets the performance requirements of multiple indicators. Integrating different neural networks in antenna design is also a topic with significant research value. In Figure 2, the first step is selecting an antenna type, which can be completed by engineers with the help of a Support Vector Machine (SVM)-based recommendation system. After the basic geometric shape of the antenna is determined, AI approaches such as artificial neural networks (ANNs) and stacked ensemble learning models can be used to calculate the values of model parameters.
3.3 Electromagnetic Simulation
In electromagnetic simulation, AI can be used to comprehensively evaluate system performance through multi-physical field coupling simulation, including electromagnetic, thermal, and structure simulation. Interdisciplinary simulation offers in-depth insights into RF and antenna design, guaranteeing advanced design robustness and adaptability. AI algorithms can be used to quickly calculate electromagnetic radiation and scattering, and extract key information from measurement data to optimize design parameters, demonstrating significant potential in the inversion analysis of electromagnetic simulation. This approach not only improves design accuracy but also accelerates the implementation of design concepts in products. Additionally, AI technologies can also be applied in real-time electromagnetic simulation to quickly obtain design results and significantly accelerate design iteration. Fast iteration is critical to expediting prototype design and testing, significantly reducing the time required for product development. Last but not least, AI plays a critical role in quantifying uncertainties that may affect design reliability in electromagnetic simulation. Such uncertainties can be evaluated using AI technologies to provide a more reliable basis for decision-making in the design process, thereby improving the design success rate.
In forward inversion calculation between the source and the scattering field, matrix inversion needs to be performed during conventional full-wave simulation of electromagnetic scattering and radiation to calculate the induced current. This time-consuming process can be expressed as:
\(\bar{J}=\left ( \overset{=}{I} - \overset{=}{\chi } \cdot \overset{=}{G }_{D} \right ) ^{-1} \cdot \left ( \overset{=}{\chi } \cdot \bar{E}_{inc} \right ) \) (1)
To accelerate this process, researchers have proposed many DL-based non-iterative approaches and found that GAN-based AI approaches yield better performance than other neural networks (such as U-Net), especially when complex scatterers are involved. A prime example is the forward-induced current learning method (FICLM). It calculates induced currents through neural network mapping first and then calculates the scattering field by multiplying the Green function by the induced current. The accuracy of the calculated scattering field increases with the number of input schemes. As illustrated in Figure 3, the input scheme includes many combinations of the incident field and the background corresponding to dielectric constants. This meets the requirement of solving the electromagnetic scattering problem in wave physics.

Figure 3 FICLM flowchart
In inversion and optimization, AI models and algorithms can be used to perform inversion analysis on electromagnetic phenomena to extract background information from electromagnetic measurement data and optimize device design parameters. Three types of neural network solvers were developed to solve electromagnetic inversion problems. Figure 4 shows the design flowchart of the first type, which directly inverts the physical parameters of the scatterer from the electromagnetic measurement result. The following model expresses the learning process:
\(R_{l}=\min _{R_{\theta}, \theta} \sum_{m=1}^{M} f\left(R_{\theta}\left(\bar{E}_{m}^{s}\right), \overset{=}{\chi } _{m}\right)+g(\theta)\) (2)
We fit each pair of physical parameters \( \overset{=}{\chi } _{m} \) and scattering field \(\bar{E}_{m}^{s} \) in the training data to obtain the neural network (R_{l}\), which directly maps the scattering field to the physical parameters. We introduced a regularization term \(g(θ)\) to avoid overfitting. However, this model has limited inversion capabilities because it learns a large amount of unnecessary information (known information about wave physics) .

Figure 4 Flowchart of the first type

Figure 5 Flowchart of the third type
The second type still uses the conventional objective function approach, where a neural network is trained as a component to learn iterative solvers. The third type integrates an approximation solver (such as a back propagation) with a DNN. It converts the input \(\bar{E}_{m}^{s} \) of the DNN from a measured electromagnetic field to an approximation solution \( \overset{=}{\chi } _{m}^{s} \) of a physical parameter (see the model flowchart in Figure 5), thereby reducing the learning workload of the neural network and simplifying the learning process. Inspired by the similarity between the iterative approach of inverse scattering problems (ISPs) and the DNN architecture in the third type of neural network solvers, researchers improved the DNN topology by incorporating three CNN modules that are cascaded and trained separately. Initially, researchers focused on the contrast information of the third type. The approximate and actual results of contrast information are used as the input and output of the CNN modules. Later on, inspired by wave physics, they developed an enhanced neural network. This network introduces the induced current and electric field in inputs and outputs to significantly advance inversion performance, as shown in Figure 6. The following model expresses the learning process of the neural network that integrates electromagnetic physics:
\( R_{l}=\min_{R_{\theta },\theta } \sum_{m=1}^{M}f\left ( R_{\theta } \left ( \bar{J}^{+}, \bar{E}^{+} \right ) , \bar{J}^{l1},\bar{J}^{l2},..., \bar{J} \right )+g(\theta ) \) (3)
where \(\bar{J}^{l1},\bar{J}^{l2},..., \bar{J} \) indicates the density of the output induced current of each CNN module.

Figure 6 Electromagetic inversion flowchart for neural networks
4 Advantages of AI Technologies in RF and Antenna Design
4.1 Improved Design Efficiency
AI technologies have significantly accelerated and innovated RF and antenna design by automating the design process, including repetitive tasks such as parameter calculation and model generation. The fast simulation and optimization capabilities of AI algorithms enable designers to formulate high-quality design solutions efficiently. The emergence of intelligent design tools further enhances design efficiency and visualize the design process.
AI surrogate models significantly improve design efficiency by using ML algorithms and models to replace complex RF/microwave components. This approach achieves the same accuracy as electromagnetic models do and makes circuit models more computationally efficient by replacing the time-consuming full-wave electromagnetic simulation. However, it requires fully trained AI/ML models and a certain amount of data to do so. Nevertheless, this approach not only improves design efficiency but also reduces the development cost, accelerating the transformation from concept to implementation in RF and antenna design.
4.2 Optimal Design and Innovation
AI offers advanced optimization capabilities for RF and antenna design. For instance, a multi-objective AI model can take many design objectives into account, such as performance, cost, and size, to find the optimal design. AI also offers advanced data analysis capabilities for fine-grained control of design details, achieving global optimization and further enhancing design quality. Furthermore, AI provides innovative design capabilities to inspire novel design ideas and approaches.
In the design optimization of complex RF devices, AI models trained on numerical and experimental data offer excellent optimization capabilities, which can be used to design complex antennas, arrays, and RF devices — a task difficult for conventional design approaches. Designing complex RF devices is equivalent to solving high-dimensional and non-convex optimization problems and usually involves time, frequency, and spectral domains and many constraints. Such problems are challenging due to the characteristics of nonlinearity, multiple scales, and strong mutual coupling. Optimization algorithms often require numerous simulation results to achieve the performance goal, and the number of iterations depends on which optimization algorithm is used and the complexity of the problem. Full-wave electromagnetic simulation is a suboptimal approach because it is both energy and computationally inefficient, especially for fast simulation and optimization of electrically large electromagnetic models. ML provides an advanced solution to these problems by achieving the same precision as an electromagnetic model does without time-consuming simulation, significantly improving the efficiency and feasibility of optimization.
4.3 Able to Solve Complex Problems
AI technologies can be used to find optimal solutions for RF and antenna design, especially for design scenarios with high dimensions and complex constraints. AI algorithms excel in identifying the characteristics and behavior of non-linear problems and provide more accurate analysis results, which conventional approaches are incapable of.
AI also excels in analyzing multi-physical field coupling by taking different information about physical fields into account, such as electromagnetic, thermal, and structure information, comprehensively evaluating the performance of RF and antenna systems. This interdisciplinary analysis approach provides more in-depth design insights to ensure optimal system performance from various perspectives. Additionally, AI technologies deliver significant value in managing uncertainties, such as ever-changing material parameters and working environments, achieving a robust design and assuring high stability and reliability for RF and antenna systems under various conditions. AI has become indispensable in RF and antenna design due to its advanced capabilities, facilitating innovation in design approaches and improving design quality.
In model analysis of complex RF/microwave structures, two representation approaches are often used: the electromagnetic model (fine model) and the lumped parameter equivalent circuit model (rough model). The fine model is more accurate but is computationally more expensive. The rough model is computationally efficient but inaccurate. Traditional spatial mapping technology optimizes geometric parameters and achieves performance goals by converting parameters between these models. This parameter conversion process can be completed more efficiently by AI/ML models in fewer computational steps. These models can identify the complex relationships between the fine and rough models by learning a large amount of data to quickly and accurately adjust parameters in the optimization process. This approach requires fewer computing resources and accelerates design iteration, enabling designers to design high-performance RF/ microwave structures more efficiently in a cost-effective manner.
4.4 Adaptive and Dynamic Adjustment
Adaptive and dynamic adjustment is key to significantly advancing the applicability and robustness of RF and antenna systems. Multi-function AI systems can be deployed to monitor and adjust RF and antenna systems in real time, significantly improving system performance from various perspectives, as shown in Figure 7.
These AI systems can monitor system parameters in real time and automatically adjust RF circuit and antenna parameters through intelligent algorithms under ever-changing working conditions, guaranteeing the performance and stability of the communications system. Such real-time monitoring and adjustment capabilities significantly enhance the flexibility and response speed of RF circuits and antenna systems.

Figure 7 AI-based adaptive adjustment in RF circuit and antenna design
AI technologies also play a significant role in intelligent fault diagnosis by analyzing system running data to detect faults and forecast problems quickly. Based on the analysis result, preventive or corrective measures can be taken to improve system availability and reliability. This approach reduces the labor and time costs of system maintenance and prevents service interruptions caused by faults.
Furthermore, AI technologies enable RF and antenna systems to better adapt to dynamic environments, for instance, mobile communications scenarios, where signal interference and other environmental factors may affect communication quality. Such environmental changes can be analyzed in real time using AI to automatically adjust the beam status of antenna arrays, optimize signal transmission and reception performance, and ensure continuous and stable communication.
4.5 Interdisciplinary Integration and Personalization
One of the advantages of AI technologies in circuit and antenna design is promoting interdisciplinary integration and personalized design. AI can be used to integrate advanced technologies and concepts from different fields, such as material science, electronic engineering, and computer science, to achieve converged innovation, while offering customized RF and antenna design solutions based on specific requirements and application scenarios.
Interdisciplinary integration creates new perspectives and solutions to RF circuit and antenna design, facilitating technological innovation and development.
In personalized design, the application of AI technologies improves the flexibility and degree of customization of the design process. AI can be used to generate customized design solutions by analyzing the requirements of specific applications, such as the communication range, frequency bandwidth, and environmental factors. This capability meets a diverse range of market requirements and offers optimal solutions for specific applications, enhancing the performance and applicability of RF systems.
5 AI Model Design Examples
As AI technologies evolve, they demonstrate significant advantages in RF circuit and antenna design, one of the most crucial and challenging tasks in wireless communications systems. In this section, we discuss several filter design cases to highlight the application of AI in RF and antenna design.
5.1 Bandpass and High-Pass Filters
The topology of a device can be pixelated to achieve a high degree of freedom for device variables and improve the odds of achieving the performance goal. This approach enjoys widespread attention from the microwave research community due to its advanced performance. It can be used to implement all kinds of microwave circuits, including power amplifiers and antennas. Based on the idea of pixelation, we designed different types of microwave filters that can operate on the Ka frequency band (26.5 GHz to 40 GHz) and meet specific frequency selectivity requirements. These filters have been widely used in autonomous driving vehicles. The first example includes bandpass and high-pass filters. We used alumina with a thickness of 0.127 mm as the substrate and gold with a thickness of 4 μm as transmission cables to design the filters. The design process is as follows: (1) Create an electromagnetic model of a solid patch on commercial software such as CST and HFSS. The patch has two ports with 50 ohm microstrips. (2) Divide the patch into grids of different sizes based on the quantity of resonators. (3) Set each grid to metal (1) or non-metal (0) for the neural network's training process. (4) Configure the Python environment to run the electromagnetic software automatically. (5) Generate data using the electromagnetic software. (6) Train a CNN model using binary sequences (1: metal; 0: non-metal) as the input and \(S_{11}\) and \(S_{21}\) parameters as the output. (7) Replace the electromagnetic model with the trained CNN model, which is computationally more efficient. (8) Develop an optimization algorithm based on the CNN model. The algorithm uses the passband return loss and stopband insertion loss as the input to generate the required filter shape.

Figure 8 Simulation diagram of a pixelated filter
Figure 8 illustrates the primary filter structure configured on CST. To accelerate data generation, we used a symmetrical plane in the target structure, where the right half of the structure is a mirror of the left half. We divided the left half plane into 20 rectangular grids (4 x 5). Based on the experience we amassed from coupling design, we set some of the grids to 1 (red) or 0 (black) before the optimization, as shown in Figure 8. We also designed different grid sizes for the input/output (I/O), inter-resonator (I/R) coupling, and resonator areas to accelerate full-wave simulation without compromising model precision. The size of each rectangular grid in the I/O and coupling areas is 0.12 mm x 0.3 mm. The size of each grid in the resonator area is 0.27 mm x 0.3 mm. Because the coupling area (coupling gap) is too small compared to the resonator size in most designs, we chose a smaller grid size for the potential coupling area and a larger grid size for the resonator area to reduce the total number of variables in the target design space. In our design, 15 grids are considered variables. We found 32,768 possible structures in total by specifying metal and non-metal grids, performed 32,768 full-wave electromagnetic simulations on frequencies ranging from 34 GHz to 46 GHz, and obtained 51 samples. We used 15 binary variables and frequency variables as the input, and the real and imaginary parts of \(S_{11}\) and \(S_{21}\) as the output to train an EM-CNN model that can replace full-wave simulation.
Additionally, we integrated an optimization solution based on a genetic algorithm (GA) to develop filters with different filtering performance by modifying the layout and combination (distributions) of metal and non-metal grids (distributions). The stopbands of the bandpass filter are defined as follows:
- Stopband 1 (SB-1) ≤ 40 GHz
- Stopband 2 (SB-2) ≥ 44 GHz
- Passband: 41 GHz ≤ Passband (PB) ≤ 43 GHz
- Passband insertion loss: \(max[S_{21}(PB)] > −1 dB\)
- Maximum passband return loss: \(max[S_{11}(PB)] < −10 dB \)
- Stopband suppression: \(max[S_{21}(SB)] < −25 dB\)
The stopbands of the high-pass filter are defined as follows:
- Stopband 1 (SB-1) ≤ 40 GHz
- Passband: ≥ 42 GHz
- Minimum in-band insertion loss: \(max[S_{21}(PB)] > −1 dB\)
- Maximum passband return loss: \(max[S_{11}(PB)] < −10 dB\)
- Stopband suppression: \(max[S_{21}(SB)] < −25 dB \)
Based on these design objectives, the loss function of the filters can be expressed as:
\( K=max\left [ (S_{11})_{PB},- RL \right ] + w*max\left [ (S_{21})_{SB},- IL \right ] \) (4)
where w indicates the weighting factor between the passband return loss (RL ) and the stopband insertion loss (IL ). The weighting factor w is a key factor for adjusting the passband return loss and the stopband insertion loss. The factor is determined according to the sensitivity of the optimization parameters and their collective impacts on the return loss and insertion loss in the passband and stopband. The EM-CNN model is then used to optimize the loss function for the bandpass filter, generating a binary sequence [1, 0, 0, 0, 1, 0, 1, 0, 1, 0, 0, 1, 1, 1, 1], which is converted into a geometric shape on CST, as shown in Figure 9a. The binary sequence generated for the high-pass filter is [0, 1, 0, 1, 1, 1, 0, 0, 1, 1, 1, 1, 1, 1, 1], which is also converted to a geometric shape on CST, as shown in Figure 9b. Figure 10a and Figure 10b illustrate the full-wave simulation performance of these filters, which deliver optimal in-band and out-of-band performance that meet the requirements of our design.

Figure 9 Optimization results of pixelated filters

Figure 10 Simulation results of optimized filters
5.2 Dual-Passband Filter
We also used the EM-CNN model to design a dual-passband filter, which requires a more complex objective function than single-passband filters.
- Stopband 1: 36 GHz ≤ Stopband (SB-1) ≤ 38 GHz
- Stopband 2: 44 GHz ≤ Stopband (SB-2) ≤ 46 GHz
- Passband 1: 34 GHz ≤ Passband (PB1) ≤ 36 GHz
- Passband 2: 38 GHz ≤ Passband (PB2) ≤ 42 GHz
- Passband insertion loss: \( max[S_{21}(PB)] > –2 dB \)
- Maximum in-band return loss: \( max[S_{11}(PB)] < –15 dB \)
- Out-of-band rejection: \( max[S_{21}(SB)] < –25 dB \)
We set w to 0.8 to balance the passband return loss and stopband insertion loss, minimizing the loss. Similar to the previous examples, we used a GA-based optimization algorithm to obtain a binary sequence [1, 0, 1, 0, 0, 1, 1, 1, 1, 0, 1, 1, 1, 1, 0, 1] that characterizes the geometric shape of the filter and the corresponding target response. Figure 9c shows the geometric shape of the filter corresponding to the optimized binary sequence. We compared the CNN prediction results with the full-wave simulation result. Figure 11a shows the S parameter response of the filter predicted by the CNN model. Figure 11b shows the S parameter response calculated by CST. We can see that the response calculated by CST is highly similar to that predicted by the CNN model and fulfills the filter performance goal. This example is only used to explain the feasibility of the design, which does not meet all performance requirements due to a short design time. Increasing the degree of freedom and the density of grids can further improve filter performance.

Figure 11 Reponse comparison of optimized filters
6 Challenges and Prospects
6.1 Data Quality and Availability
The generation, quality, and availability of data are critical to the successful application of AI models in antenna design. Large-scale and high-quality data is the cornerstone of AI model training. However, RF and antenna design faces several data challenges: lack of data, heavy reliance on expertise, and high complexity. Data used in RF and antenna design often comes from electromagnetic model simulation, which is computationally expensive and makes data obtaining more difficult.
An accurate AI/ML model needs to be trained on adequate training, validation, and test datasets that fully characterize the problem to be solved. For RF/microwave structures, these datasets are usually generated through electromagnetic model simulation, which is an expensive and time-consuming process. Additionally, improving the accuracy and generalization capability of a model involves data collection, organization, and labeling. This process requires careful selection and sampling of data to ensure that the quality and availability of the datasets can characterize the key areas in the design space.
To address the data challenges faced by AI model training in RF and antenna design, we can implement various policies to strike a balance between data generation time and model accuracy: (1) Active learning: Selectively generates the most informational data points, reducing the total amount of required data while maintaining high model accuracy. (2) Transfer learning: Creates models based on pre-trained models that have learned features and patterns from related tasks, significantly reducing the amount of new data required for model training. (3) Data augmentation: Increases the scale of datasets, reduces the demand for extra electromagnetic simulation, and maintains the diversity of training data. (4) Dimensionality reduction: Focuses on the most crucial parameters, simplifies the data generation process, and reduces the computational cost. A strategic trade-off is required between minimum data generation time and maximum model accuracy.
In addition to data collection and organization, data can be generated using AI technologies. For instance, diffusion models and other generative models offer an innovative data generation approach by learning the distribution features of current data and generating new data samples that are similar to the current data. These models can be used to generate more training data from limited measurement or simulation data for RF and antenna design. These modes can also identify complex data structures, improving the diversity of the generated data while maintaining consistency with the original dataset. GAN is also an advanced data generation approach that can generate realistic data samples through adversarial training. In RF and antenna design, GANs can be used to generate electromagnetic simulation data to help designers evaluate the performance of different design solutions in the early design phase. In conclusion, the policies mentioned above can be used to reach an optimal trade-off between minimum data generation time and maximum model accuracy and develop effective AI/ML models for RF/microwave structures, accelerating the development process of AI models and offering innovative and efficient RF and antenna design solutions.
6.2 Model Selection and Parameter Configuration
In RF/microwave structure design, AI/ML model development faces a daunting challenge: model selection and hyperparameter configuration. Because different problems have different characteristics, no model can meet the requirements of all scenarios. We need to select the most suitable model architecture based on the problem description. Additionally, selecting hyperparameters, such as the number of layers, number of neurons, data segmentation ratio, and activation function, has a significant impact on model performance. Current hyperparameter selection approaches heavily depend on expertise and trial-and-error tests, which are laborious and time-consuming, and may cause uncertainties in the model development process. Efficiently and accurately selecting models and configuring hyperparameters are the essential challenges for the development of high-performance AI/ML models.
To address these challenges, we can implement the following policies: (1) Develop automated model selection tools that can recommend or select the most suitable model architecture based on the characteristics and data features of the problem. (2) Use hyperparameter optimization techniques, such as grid search, random search, or Bayesian optimization, to systematically explore the hyperparameter space and find the optimal hyperparameter combination. (3) Develop an automated development process integrating data pre-processing, model training, hyperparameter optimization, and model evaluation to reduce manual intervention and improve development efficiency.
6.3 Algorithm Complexity and Computing Resources
Algorithm complexity and computing resources are the major obstacles to the application of AL/ML algorithms in RF and antenna design. Advanced AI algorithms, especially DL models, often require a large number of computing resources for training and inference. These resources include but are not limited to high-performance GPUs, large storage spaces, and fast data processing capabilities, which are expensive and difficult to obtain, especially for research teams and small development teams. Additionally, highly complex algorithms may increase the training time, slowing down design iterations and innovation processes. Balancing algorithm complexity and computing resources has become a key issue for efficient AI-assisted design.
To address these challenges, we can implement the following solutions: (1) Algorithm optimization: More efficient algorithms can be developed to reduce computing steps and resource consumption while maintaining or even improving model performance. (2) Model simplification: Technologies such as model pruning and knowledge distillation can be used to simplify the model structure and reduce model parameters, thereby lowering the computing workload. (3) Hardware acceleration: Exclusive hardware, such as tensor processing units (TPUs) and field programmable gate arrays (FPGAs), can be used to accelerate AI algorithms, offering optimal computing capabilities for DL. (4) Cloud computing resources: The advanced computing resources provided by cloud computing services can be allocated on demand to reduce investment costs for local hardware. (5) Parallel computing: Parallel computing technology can be used to allocate training tasks to different processors or devices to reduce the training time. (6) Resource scheduling and management: An intelligent resource scheduling system can be developed to optimize the allocation and use of computing resources, and improve resource efficiency. (7) Lightweight models: Lightweight AI models, such as MobileNet and ShuffleNet, can be developed for environments with limited resources. (8) Model quantization: Model quantization technology can be used to lower the requirements on model precision, memory, and storage and reduce computational complexity. (9) Heterogeneous computing resources: Different types of computing resources, such as CPUs, GPUs, and application-specific integrated circuits (ASICs), can be integrated to achieve optimal resource allocation for computing tasks. These solutions can be employed to reduce the complexity of AI algorithms, accelerate the design process, and improve the efficiency and feasibility of RF and antenna design when computing resources are limited.
6.4 Explainability and Transferability
The explainability of AI models is an important issue that is often neglected in RF and antenna design. Although some multi-objective, multi-functional AI models, especially DL models, deliver excellent performance in solving certain complex problems, they are often considered "black boxes" because it is difficult to understand how these models make decisions. In RF and antenna design, designers need to understand the decision-making process of the model to ensure that the design meets the requirements of the principles of physics and the application scenarios. Lack of explainability not only increases design risks, but also limits the application of AI models in key application scenarios.
To improve model explainability in RF and antenna design, we can adopt the following approaches: (1) Develop and apply explainable AI technologies, such as Local Interpretable Model-Agnostic Explanations (LIME) and SHAPley Additive exPlanations (SHAP), which can explain model prediction results. (2) Use visualization tools to demonstrate the working principles of models, including the significance of features and decision boundaries. (3) Use simpler and easy-to-understand models, such as decision trees or linear models, although they may underperform complex models in certain scenarios. (4) Use model distillation technology to migrate the knowledge of complex models to a simpler model. (5) Implement post-processing technologies, such as rule extraction, to extract explainable rules from black-box models. (6) Take explainability into account in the model design phase and select naturally transparent model architectures. These approaches can improve the explainability of AI models in RF and antenna design, enhance designers' trust in models, and promote the application of AI technologies in this field.
However, there is a trade-off between model explainability and transferability. For instance, complex models may yield good performance in new environments (good transferability), but they may be difficult to explain (suboptimal explainability). Simple models are easy to explain but they may deliver unsatisfactory performance in diversified environments or scenarios with a diverse range of data (suboptimal robustness and transferability). To compensate for the lack of transferability, multiple models must be designed for different scenarios in the early design phase, increasing the design cost and maintenance difficulty. These models may fail to adapt to ever-changing application scenarios. Consequently, new and customized AI models need to be developed for RF circuit networks, just as the world of physics needs a general theory. These models must have high explainability and advanced generalization capability (transferability). Developing such models is one of the key challenges for the extensive application of AI technologies in RF and antenna design.
6.5 Multi-Domain Convergence
RF and antenna design is a cross-disciplinary field involving many disciplines, such as electromagnetics, material science, and electronic engineering. The application of AI models in this field requires the processing and integration of complex data and knowledge from different disciplines. However, multi-domain convergence faces many challenges in AI model development. Data from different disciplines may have different characteristics and formats, making it difficult to integrate. Additionally, expertise and theories in different domains must be integrated to ensure that AI models can thoroughly understand and solve problems. Current AI models often focus on only one domain and lack interdisciplinary integration and collaboration, limiting their performance and application in RF and antenna design.
To address these challenges, we can implement the following approaches: (1) Set up an AI model development team of experts from different domains to ensure that multi-disciplinary knowledge and data are considered. (2) Develop a data standardization process to unify the formats of data collected from different sources, facilitating data processing and analysis by AI models. (3) Use multi-task learning technology to enable AI models to simultaneously learn multiple tasks, promoting the convergence of knowledge from different domains. (4) Implement transfer learning technology to migrate the knowledge learned in one domain to another, improving the adaptability and performance of models in the new domain. (5) Design domain-specific model architectures that can better characterize and process data and knowledge from specific disciplines. (6) Develop cross-disciplinary knowledge graphs to integrate professional knowledge in different domains and provide rich background knowledge for AI models.
7 Conclusion
In this paper, we discussed in detail the application, advantages, challenges, and future development of AI in RF and antenna design. AI technologies have demonstrated significant potential in antenna optimization, RF circuit design, and electromagnetic simulation by improving design efficiency, optimizing design results, and solving complex problems. Although AI faces many challenges in RF and antenna design, such as data quality, model selection, algorithm complexity, model explainability, and model transferability, AI still enjoys a promising prospect in this field. We believe that the continuous research and innovation on AI will facilitate more significant breakthroughs and progress in RF and antenna design. AI has opened up a new horizon for RF and antenna design by improving design efficiency and optimizing performance, facilitating the development of wireless communications. However, certain technological and data challenges need to be addressed before AI can fully unlock its potential in the field. Future research is expected to focus on improving data quality, developing more efficient algorithms, enhancing model explainability, and promoting multi-domain convergence. Such research will advance the development of wireless communications systems and deliver significant benefits to our daily lives.