Design and Optimization of Antenna for Next Generation

Abstract

The rapid evolution of fifth-generation (5G) and beyond fifth-generation (B5G) wireless communication systems has significantly increased the demand for compact, high-performance, and computationally efficient millimeter-wave (mmWave) antenna architectures. However, the design of mmWave Multiple-Input Multiple-Output (MIMO) antennas remains highly challenging due to the simultaneous requirements of high gain, wide bandwidth, compact in- tegration, low mutual coupling, and reduced computational complexity. In addition, conven- tional full-wave electromagnetic optimization techniques are computationally expensive and time-consuming, particularly for iterative antenna-design procedures. To address these chal- lenges, this doctoral thesis proposes a hybrid electromagnetic and physics-guided artificial intelligence framework for the design, optimization, and rapid synthesis of high-performance mmWave MIMO antennas. The proposed methodology combines advanced electromagnetic engineering techniques with physics-guided neural learning strategies to improve both an- tenna performance and computational efficiency. In the first part of this work, a compact four-element 28 GHz MIMO antenna system was designed, fabricated, and experimentally validated for 5G mmWave applications. The proposed antenna architecture integrates De- fected Ground Structures (DGS) and a Frequency Selective Surface (FSS) superstrate to re- duce mutual coupling and enhance radiation performance. The incorporation of the FSS su- perstrate at an optimized distance of 7 mm improved the antenna gain from 4.8 dBi to 8.4 dBi, corresponding to an enhancement of approximately 75% (+3.6 dBi). In addition, the proposed decoupling configuration achieved port-to-port isolation exceeding 25 dB while maintaining satisfactory MIMO diversity performance. To overcome the computational limitations of conventional full-wave solvers, a dual-network Physics-guided AI framework was developed for inverse antenna synthesis and forward electromagnetic prediction. The inverse synthesis model was trained using 8,000 analytically generated samples derived from Transmission- Line Model (TLM) equations, while the forward analysis model was independently trained using 475 high-fidelity CST full-wave simulation samples. By integrating electromagnetic prior knowledge into the neural learning process, the proposed framework preserves physical consistency while improving prediction accuracy and computational efficiency. The devel- oped framework reduced the antenna synthesis time from several hours of full-wave sim- ulations to only a few milliseconds. The inverse synthesis model achieved sub-millimeter prediction accuracy with a Root Mean Squared Error (RMSE) of Wp (≈ 0.00156 mm) and Lp (≈ 0.00169 mm) and a regression coefficient of (R2 = 0.99999), while the forward predic- tion model limited the relative resonant frequency prediction error to only 0.08% compared with CST full-wave simulation results. Overall, the proposed hybrid electromagnetic–AI framework establishes a scalable and computationally efficient methodology for intelligent antenna synthesis and multi-objective optimization in next-generation wireless communica- tion systems, with strong applicability to future 5G-Advanced and 6G mmWave technologies.