In the modern era, antennas are crucial in the arena of wireless communication. Various applications like satellite communication systems, unmanned aircraft, radar systems, WLAN, WiMAX, missiles, and wireless telecommunications are constrained by size, cost, weight, ease of production, and ease of implementation. In microwave applications like satellite communications and maritime radar communications, dielectric resonator antennas can perform very well for narrowband, wideband, and ultra-wideband (UWB). UWB technology has gained a significant boost, as it is the unlicensed frequency band from 3.1 GHZ to 10.6 GHz with a continuous bandwidth of 7.5 GHz for commercial communication applications [1].

The most notable attributes that enhance the applicability of dielectric resonator (DR) antennas are wide data transmission capability, lightweight, low metallic losses, and high radiation efficiency [2,3]. In DRA, various strategies have been proposed in the literature to improve DRA performance, such as modifying the shape of the DR, incorporating metallic strips, using cross-slot coupling, and employing single or dual-feed configurations. In particular, circularly polarized (CP) DRAs can reduce polarisation mismatch and minimize multipath interference [4,5], thereby enhancing anti-interference capabilities and ensuring high-quality electromagnetic signal transmission. Therefore, satellite communication, navigation, and global positioning systems all have strong application potential. CP antennas have a significant role in the advancement of high-speed wireless networks, due to their effectiveness in suppressing polarization mismatches and mitigating multipath effects. Broadband CP antennas are especially desirable for wireless systems requiring high data rates; however, traditional microstrip antennas often suffer from limitations such as narrow bandwidth and high axial ratios. The increasing demand for compact CP antennas with wide bandwidth has positioned them as a critical research focus, particularly for 5G, wireless, and satellite communication systems. Their ability to overcome multipath losses and polarization mismatches underscores their importance in modern communication technologies.

Therefore, research on circularly polarized DRAs has recently attracted more interest [[5], [6], [7], [8], [9], [10]]. Various methods have been presented in the literature to achieve circular polarization [[11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23]]. The DRA is excited by a single feed to obtain circular polarization through a structure-oriented DR. The shape and dimensions of the dielectric resonator are altered to widen the axial ratio bandwidth. In [11], orthogonal modes are excited by adding metallic rings and vertical slots in the circular waveguide, enhancing the Eθ and Eϕ components. A 90-degree phase-shifted hybrid resonator is used [12] to excite circular polarization, and the axial ratio bandwidth is enhanced by embedding a dielectric resonator in the planar antenna [13]. Etching diagonal slots [14], adding a cross-dipole structure to generate a quasi 90° phase delay [15], modifying the ground plane by incorporating a Jerusalem cross with two horizontal stubs [16], integrating a reconfigurable power divider [17], using a non-radiating dielectric polarizer with an aperture-coupled antenna [18,19], rotating a rectangular patch antenna by 45° and adding swastika-shaped slots in an aperture-coupled configuration [20], introducing multiple slots in the ground plane [21], integrating an asymmetric triple-dipole [22], and employing a meandered dipole in parallel with Yagi-Uda dipole elements [23] are among the recently implemented approaches for achieving circular polarization.

Predicting the axial ratio (AR) of a dual-band circularly polarized DRA remains a significant challenge due to the complex relationship between the antenna’s geometric parameters and polarization characteristics. Traditional design approaches rely heavily on computationally intensive simulations and trial-and-error methods, making them time-consuming and resource-demanding. Machine learning (ML) offers a powerful alternative by learning complex patterns from simulation data and accurately predicting the axial ratio (AR) at specific frequencies, thereby reducing design time and enhancing optimization efficiency. To overcome these challenges, the use of ML-based models is increasingly adopted for efficient antenna performance estimation. Various ML techniques have been applied to antenna design and optimization. For example, Sharma et al. employed Lasso regression, KNN, and ANN to enhance the performance of a dual T-shaped patch antenna. Semi-supervised Gaussian Process Regression (GPR) was used in [19] to predict the resonance frequency of a microstrip patch antenna. The notch frequency of a slotted patch antenna was predicted using a multi-adaptive neuro-fuzzy model [20]. Additionally, different ML models have been utilized to optimize CPW-fed radiators, producing predictions that align well with both simulation and experimental results [[24], [25], [26], [27], [28], [29], [30], [31], [32], [33], [34], [35]].

The Extra Trees regression model has shown promising results in predicting antenna performance. For example, in [36], it was used to predict the gain of a MIMO antenna. Similarly, impedance parameters have been predicted using Random Forest and Ridge Regression techniques [37,38]. Resonance frequency prediction using Random Forest Regression has also been demonstrated in [[39], [40], [41], [42]]. Recent research indicates that various antenna parameters—such as frequency, gain, and impedance—for different antenna types, including microstrip patch, Yagi-Uda, RFID, DRA, and PIFA, can be effectively predicted using ML models. Integrating ML techniques into antenna design and tuning processes enables real-time optimization and performance enhancement. The major contributions and innovations of this study are summarized as follows:

• •
  An asymmetrical dielectric resonator antenna (DRA) structure composed of three ceramic blocks is designed to achieve dual-band circular polarization covering the C-band and X-band.
• •
  A wide 10 dB impedance bandwidth of 79 % (5.2–12 GHz) and dual 3-dB axial ratio bandwidths (6.0–8.7 GHz and 9.9–11.8 GHz) are achieved, with a stable gain exceeding 4 dBi.
• •
  A machine learning (ML)-based supervised regression approach is implemented to predict the axial ratio at 7.2 GHz and 10.5 GHz.
• •
  Among various tested regression models, the Extra Tree Regression model achieves the highest prediction accuracy with the lowest error.

The paper is organized as follows: Section 2 presents the antenna design analysis. Section 3 discusses field mode excitation and circular polarization. Section 4 provides antenna performance results and discussion. Section 5 gives the array configuration analysis. Section 6 explains machine learning modeling for axial ratio prediction. Section 7 details the prediction using the Extra Tree Regression model. Section 8 concludes the paper.