I. INTRODUCTION

As we enter the next-G era, the number of antennas is increasing rapidly, leading to a growing concern regarding the coexistence and miniaturization of variable antennas. As a result, multi-port antennas have gained more and more attention as they offer more functionality or flexibility with a compact appearance [1]. On the other hand, the challenge of antenna decoupling in MIMO and full-duplex systems has emerged as a prominent area of research focus [2]- [4]. For example, in portable antenna design, imposed by the size limitations of mobile terminals, two primary objectives are consistently pursued. One involves maximizing the placement of patch antennas, while the other entails achieving high isolation and low correlations among distinct antennas [5]. However, the growing complexity of system integration in specific small areas makes it increasingly challenging to attain the aforementioned goals through conventional antenna design methods with optimized template-based geometries. Therefore, developing ML-based methods for passive design has become more and more attractive in recent years [6]- [8]. In this study, we investigate the deep learning (DL) based inverse design of multi-port antennas to enable dual-band and MIMO operations.

II. DL-BASED MODELING AND INVERSE DESIGN

The main component of the inverse design approach is the CNN-based EM predictor. As shown in Fig. 1, the input of the CNN is a pixelated multi-port antenna structure divided into 12 × 12 pixels. Antenna ports are placed randomly to the left and right halves. Amplitudes of S 11 , S 22 and S 21 were converted to dB and clipped between -20 and 0 for the training. The output of the CNN covers the frequency range of 20-40 GHz with 81 linearly placed points. As 3 terms are predicted at a given frequency, the output has 243 neurons. For the preceding layers, hyper-parameters are decided in accordance to [6]- [8].

It is worth noting that a two-step training approach was adopted for efficient use of computational resources [8]. A total of 300K fast simulations and 128K accurate simulations were conducted. These datasets were augmented with geometrical transformations. In the first step of training, a randomly initialized network was trained with a test-validation-training split of 38K-38K-375K over 30 epochs. The second step essentially carries on the training of the first model for another 30 epochs with the dataset of accurate simulations. This set of data is divided into 15K-18K-175K test-validation-training splits. Once the training was complete, we deployed the network with a genetic algorithm (GA) to optimize for different antenna properties. With the help of a GPU, prediction of S-Parameters for the population size of ≈4000 structures takes less than 1 second, and optimization concludes within ≈2 minutes for 100 generations.

III. EXAMPLE DESIGNS

To demonstrate the effectiveness of the modeling approach and the inverse design methodology, we present synthesis results antennas and their measured properties in Fig. 2. Fig. 2a shows an example a dual-band 2-port patch antenna on a 12 mil RO4003C substrate operating at 28.3 and 30.3 GHz. It should be noted that the goal here is to match S 11 and S 22 at 2 distinct frequencies while ensuring isolation (S 21 ). Furthermore, it was aimed that S 11 and S 22 should be mismatched outside of the target band. The resulting EM structure implements these requirements. Fig. 2-b implements similar functions at a slightly different frequency range of 28.3 and 29.3 GHz. Fig. 2-c shows the simultaneous matching of 2 antenna ports while providing more than 15 dB isolation at 27.5 GHz. These examples illustrate that CNN-based multiport EM modeling for antenna synthesis is a promising method for optimizing various aspects of multi-port antennas.

IV. CONCLUSION

The presented approach with deep CNN carries out an extension of the previous works by providing greater design freedoms. For example, as we allow a single antenna to be excited from multiple ports, the bandwidth of the antenna can be compositely extended. Moreover, a single radiating structure can be used for both transmit and receive chains by isolating antenna ports. While typically such functionalities require considerable engineering effort, CNN-based modeling and inverse design could synthesize a solution rapidly, as shown in example designs. In addition, once trained, CNN can be repeatedly utilized for antenna design to compensate for the initial computational investment. These aspects can make CNN-based modelling a viable design tool.