This paper proposes a data-driven design approach for generating wideband frequency-invariant uniform beams to mitigate the beam squint phenomenon, where beamwidth varies with frequency in antenna array systems. The proposed framework utilizes three machine learning models—supervised learning, unsupervised learning, and reinforcement learning—to derive optimal weight vectors that maintain a constant beamwidth across the entire operating band while maximizing reception performance. In particular, robustness against hardware imperfections and real-world channel environments is secured by pre-training the models with simulation data and subsequently fine-tuning them using calibrated data acquired from outdoor radiation tests. For performance analysis, frequency invariance is validated through the root mean square error of the half-power beamwidth with respect to a target uniform beamwidth of 20°. Furthermore, bit error rate analysis confirms that the proposed method maintains reception performance comparable to the theoretical Linearly Constrained Minimum Variance (LCMV) beamformer.
Chang et al. (Fri,) studied this question.