Machine learning in pre- and post-manufacturing stages of active array antennas

Abstract

The rapid advancement of next-generation wireless communication and sensing technologies demands high-performance, intelligent antenna systems capable of adaptive beamforming, efficient signal control, and reliable operation under dynamic conditions. Active phased arrays, especially non-uniform (aperiodic) layouts, offer flexibility for power-efficient side-lobe control and improved spatial coverage, but at the cost of increased electromagnetic, thermal, and computational complexity. Practical IC-integrated aperiodic arrays therefore require innovative design and diagnostic frameworks that efficiently address challenges across pre- and post-manufacturing stages.

Traditionally, design and performance evaluation of active phased arrays have relied on deterministic electromagnetic simulations or optimization-based methods focusing on radiation characteristics such as gain, beamwidth, and side-lobe levels (SLL). As arrays grow more complex, such methods face limitations in scalability and computational feasibility. Machine learning (ML) offers a pathway to overcome these barriers by enabling data-driven modeling, optimization, and diagnosis of large-scale or non-uniform active arrays. This interdisciplinary challenge lies at the intersection of electromagnetics (EM), array modeling, and data-driven methods, requiring frameworks that balance physical accuracy with computational efficiency.

The objective of this doctoral research is to develop and validate ML-based methodologies that enhance both the design and operation of active arrays, emphasizing efficient modeling, synthesis, and diagnosis under mutual coupling (MC), environmental effects, and fabrication-induced uncertainties. The research addresses both pre-manufacturing (e.g., synthesis, EM modeling, and topology optimization) and post-manufacturing (e.g., calibration and fault detection) aspects of active arrays, with a focus on data-efficient and physics-supported ML techniques.

Chapter II introduces the motivation, state-of-the-art challenges in phased array systems, and the potential of ML in overcoming these limitations. A detailed analysis highlights critical issues of mutual coupling, irregular routing, calibration complexity, and environmental sensitivity, establishing the need for intelligent, adaptive frameworks that integrate ML with physical modeling to enhance design reliability and post-deployment robustness.

Chapter III presents data-driven modeling of EM interactions in aperiodic phased arrays. A neural network (NN)-based framework predicts embedded element patterns (EEPs) across the visible space for non-uniform planar arrays. The cascaded NN architecture, combining a fully connected NN and a sub-pixel convolutional network, enables high-resolution pattern prediction with significant computational savings. The influence of dataset size and quality on prediction reliability is analyzed, revealing insights into data efficiency and model generalization.

Chapter IV introduces hybrid ML-physics approaches to improve prediction robustness. Two basis-function-assisted frameworks are developed using the Infinitesimal Dipole Model (IDM) and spherical harmonics. These models achieve compact, reliable EEP representation with smaller datasets while mitigating numerical instability. An ensemble method combining NN-based EEP prediction and constrained IDM achieves a 60% reduction in mean squared error (MSE) and improved prediction stability under MC effects.

Chapter V applies these ML-assisted models to system-level synthesis and diagnostics. A novel ML-driven optimization method for MIMO radar arrays integrates spherical harmonics-based EEP prediction into a particle swarm optimization (PSO) routine, enabling efficient MC-aware array topology design and minimizing maximum SLL in multi-beam configurations. ML-based post-manufacturing diagnostics are demonstrated for a 64-element active uniform array, where a fully connected NN detects faulty elements in real time using sparse far-field amplitude data.

The findings demonstrate that ML can bridge the gap between EM theory and practical array implementation. The developed frameworks offer data-efficient alternatives for MC-aware active array optimization and system diagnostics, paving the way toward intelligent, reliable, and self-adaptive active array systems for future wireless communication and sensing applications.