Experimental and Machine Learning Validation of a Modified 2.6 GHz Single-Element Switched-Beam Antenna for 5G sub-6 GHz Applications

Technical Summary: Experimental and Machine Learning Validation of a Modified 2.6 GHz Single-Element Switched-Beam Antenna

Problem Statement
In the context of 5G sub-6 GHz wireless communications, there is a critical need for compact, cost-effective, and high-performance antennas that can control radiation directionality to minimize signal interference. While single-element microstrip antennas offer advantages in size and integration, they typically lack inherent beam-steering capabilities. Traditional switched-beam solutions often rely on complex multi-element phased arrays or diode-controlled structures that increase power consumption, cost, and physical complexity. Furthermore, conventional antenna design relies heavily on computationally expensive and time-consuming full-wave electromagnetic (EM) simulations. Although Machine Learning (ML) has shown promise in predicting antenna performance, previous studies often focused solely on simulation-based prediction without comprehensive experimental validation, particularly for single-element switched-beam designs.

Methodology
The authors propose a design and validation framework for a single-element circular microstrip antenna operating at 2.6 GHz. The methodology is divided into three primary phases:

1. Antenna Design and Simulation:

   • Structure: The antenna is fabricated on an FR-4 substrate ($\epsilon_r = 4.4$, thickness = 1.6 mm). It features a modified circular patch with 56 holes distributed along its perimeter.
   • Beam Switching Mechanism: Beam steering is achieved by selectively short-circuiting specific groups of holes along the patch edge using copper wires, capacitors, and PIN diodes controlled by a microcontroller. By opening specific circuits and shorting others, the main beam is steered toward eight distinct angular directions (0°, 45°, 90°, 135°, 180°, 225°, 270°, 315°).
   • Data Generation: 150 datasets were generated using CST Studio Suite. 140 samples (112 for training, 28 for testing) were used for ML model development, while 10 independent samples served as a validation set. Input parameters included short-circuit direction, radial distance, number of short-circuit holes, and spacing between holes. Outputs were the main beam direction and reflection coefficient (return loss).

2. Machine Learning Modeling:

   • Four regression-based algorithms were implemented using Python: Gradient Boosting Regression (GBR), Lasso Regression (LR-Lasso), Linear Regression (LR), and Random Forest (RF).
   • The models were trained to predict beam direction and return loss based on the geometric input parameters. Performance was evaluated using the coefficient of determination ($R^2$) and Mean Absolute Error (MAE).

3. Experimental Validation:

   • A physical prototype was fabricated based on ML-predicted parameters. The prototype utilized copper shorting wires and PIN diodes controlled by a microcontroller to switch between the eight beam states.
   • Measurements were conducted at 2.6 GHz to verify return loss and radiation patterns, comparing experimental results against both CST simulations and ML predictions.

Key Contributions

• Single-Element Switched-Beam Design: The paper presents a compact, single-element antenna capable of eight-direction beam switching without the need for complex phased arrays, utilizing a short-circuit control mechanism along the patch edge.
• ML-Assisted Design Optimization: The study demonstrates the application of ML to predict both beam direction and return loss, significantly reducing the reliance on iterative EM simulations.
• Comprehensive Experimental Verification: Unlike prior works that focused primarily on simulation, this research includes the fabrication and measurement of a physical prototype, validating the ML predictions against real-world performance.
• Algorithm Comparison: The work systematically compares four ML algorithms, identifying Random Forest as the most effective model for this specific application.

Results

• ML Performance:
  • Beam Direction: The Random Forest (RF) model achieved perfect prediction accuracy with an $R^2$ of 1.0 and an MAE of 0. Linear Regression and Lasso Regression also achieved perfect scores, indicating a strong linear relationship between short-circuit geometry and beam direction.
  • Return Loss: The RF model demonstrated the most balanced performance for return loss prediction, achieving an $R^2$ of 0.805 and the lowest MAE of 0.21. While GBR achieved a higher $R^2$ (0.898), it exhibited a larger MAE (0.520).
• Experimental Outcomes:
  • The fabricated antenna successfully switched beams across all eight target directions at 2.6 GHz.
  • Measured return loss values ranged from 12.26 dB to 15.34 dB, with an average of 14.25 dB, confirming stable impedance matching.
  • Radiation pattern measurements generally aligned with simulations. However, specific deviations were observed: Case 4 shifted from 135° (simulated) to 165° (measured), Case 5 from 180° to 200°, and Case 8 from 315° to 300°. The maximum deviation was 30°.
  • The authors attribute these deviations to practical implementation factors not fully captured in the simulation, such as the parasitic impedance of PIN diodes, biasing circuits, solder joints, and connecting wires.

Significance and Claims
The paper claims that the Random Forest algorithm effectively enhances antenna design accuracy and reduces simulation time while maintaining strong agreement with experimental results. The study demonstrates that a single-element antenna can provide reliable directional coverage across discrete angular sectors suitable for compact 5G sub-6 GHz systems.

The authors modestly conclude that while the ML-assisted design approach is effective, the observed beam deviations highlight the challenges of translating simulation models to physical prototypes due to parasitic effects. They assert that the proposed design contributes to sustainable and energy-efficient wireless infrastructure by combining compact antenna geometry with intelligent optimization. The paper does not claim to solve all beam-steering challenges but positions this work as a validated proof-of-concept for integrating ML into the design of single-element switched-beam antennas. Future work is suggested to explore advanced deep learning models and expand the dataset size to further improve prediction accuracy for complex parameters like return loss.