Physics-informed deep learning-based adaptive beamforming for phased array weather radar.

This study introduces a beamforming framework based on physics-informed deep-learning for phased array weather radar systems, with particular emphasis on configurations involving a large number of antenna elements. The proposed approach is designed to achieve both high computational efficiency and reliable weather target detection. Physical constraints governing array behavior are explicitly embedded into the learning process by means of a tailored loss function that minimizes the mismatch between the predicted array output and a predefined reference response. To assess the effectiveness of the physics-informed deep neural network (PIDNN) beamformer, its performance is evaluated against two classical neural network architectures commonly used in the computer science domain and also represent two different learning-based alternatives: a shallow radial basis function neural network (RBFNN) and a deeper convolutional neural network (CNN). Simulation results demonstrate that the PIDNN consistently outperforms these models in terms of beamforming accuracy and computational time. Furthermore, the practical applicability of the proposed method is verified using real-world measurements acquired from the phased array weather radar operated at Osaka University (PAR@OU). The experimental results confirm that the PIDNN enables rapid and accurate reflectivity estimation while exhibiting superior clutter suppression compared with the other methods.