Silicon-Micromachined High-Gain Antennas and Beamforming Architectures for Sub-Terahertz Communication and Sensing

QC 20251006

Abstract

The increasing demands for high-speed wireless communication, intelligent sensing, and high-resolution imaging have driven interest toward the underutilized sub-terahertz (sub-THz) frequency spectrum. This region offers large bandwidths and high spatial resolution, making it a promising candidate for next-generation communication and sensing systems. However, realizing practical sub-THz systems presents several challenges, including severe path loss, low power efficiency, and significant hardware complexity due to frequency-dependent losses and fabrication constraints.

This thesis addresses these challenges by proposing a set of passive, highperformance components and system-level architectures focused on antenna design, beamforming techniques, channel modeling, and imaging methods. These components are fabricated using silicon micromachining, a scalable technology that enables the realization of compact, high-frequency passive structures with low loss and micrometer feature size. The thesis begins by developing a ray-tracing-based statistical channel model that captures essential propagation phenomena, including molecular absorption, reflection,scattering, and diffraction. The model evaluates the root-mean-square delay spread, coherence bandwidth, and subchannel stability under varying link distances, antenna gains, and misalignment scenarios. The results reveal that higher-frequency bands exhibit reduced delay spread variability and allow for more robust multi-carrier communication through channel bonding, forming the foundation for hardware-aware THz link design.

Two high-gain silicon micromachined lens antennas are introduced next. The first design is an elliptical Fresnel zone planar lens antenna that achieves circular polarization across the 500–750 GHz band. The return loss remains better than -15 dB across this range, with a measured gain of 25.7 dBi and an axial ratio below 2.5. The second design uses a circular Fresnel lens enhanced by a graded-index dielectric perforated disc, fabricated using DRIE on a silicon-on-insulator wafer. With 13 optimized Fresnel zones, this antenna achieves 35.9 dBi gain and maintains circular polarization with an axial ratio below 2.8 dB across 40% bandwidth. These antennas demonstrate state-ofthe-art performance in compact, planar form factors. Specifically, a single silicon wafer is etched on both sides using deep reactive ion etching (DRIE)and two lithographic masks, forming structures with fixed thickness and precise vertical profiles. This makes the fabrication process simple for purely dielectric-based lens antennas.

To address the limitations of wideband beamforming at sub-THz, the thesis presents a spatial-spreading approach using frequency-orthogonal passive beam steering. A multi-feed Fresnel lens system is designed to steer each frequency sub-band into a distinct spatial direction. Using four feeds and 75 GHz of total bandwidth, 16 beams are generated to cover a 32° field of view.

Experimental results show only 0.9dB steering loss, sidelobe suppression below –22dB, and a realized gain of 32.1 dBi. The lens is compact (15.8×15.8 mm, 526 μm thick) and requires only a single RF chain per feed, significantly reducing system complexity. The spatial-spreading antenna is then paired with a convolutional neural network for adaptive beam and power allocation. The CNN estimates user location using amplitude-only received signal features and dynamically assigns beam and transmit power. The system achieves up to 61% improvement in direction-of-arrival estimation accuracy, a 94% increase in data rate, and a 30% reduction in transmit power compared to static strategies.

The final chapter investigates the effect of antenna dispersion in wideband imaging. A comparison between silicon lens and metallic horn antennas reveals that the former enables higher effective bandwidth and preserves the time-domain pulse shape. Experimental results show that lens-based antennas reduce range and cross-range localization errors by up to 64% and 68%, respectively, and improve signal-to-clutter ratio by 2.7 dB. The system achieves millimeter-level resolution and resolves targets as close as 2mm in cross-range and 3mm in range.

Using this insight, a full imaging system is demonstrated by combining frequencyorthogonal beams and a time-reversal DORT algorithm. The system reconstructs images of multiple targets without mechanical scanning. Experimental reconstructions verify resolution of 0.6 mm-radius objects and accurate discrimination between targets spaced only 2mm apart, affirming the impact of dispersion-aware design for high-resolution THz imaging.

The thesis demonstrates how silicon-micromachined, high-gain antennas and passive beamforming can be effectively combined with machine learning and wideband imaging strategies to address key limitations in sub-THz systems. The proposed components are validated across communication and sensing contexts, establishing a robust framework for compact, scalable THz frontend architectures.

urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-371159