A physics-guided deep learning method for reconstructing multipolar acoustic sources

Reconstructing multipolar acoustic sources from far-field data is a highly ill-posed inverse problem when the count, locations, types, and intensities are all unknown. We propose a physics-guided deep learning framework that fully reconstructs the sources at a fixed frequency, using four-channel Direct Sampling Method (DSM) indicator functions as physics-informed inputs within a divide-and-conquer pipeline. Numerical experiments demonstrate accuracy, robustness to noise, and flexibility.