An Effective Radiation Source Reconstruction Method From Phaseless Near-Field Based on Neural Network Framework

A physical radiation source can be presented by equivalent dipoles from the source's phaseless near-field, a process known as dipole source reconstruction. Such dipole source reconstruction methods typically rely on global optimization algorithms, which pose challenges when applied to a complex source that requires a large number of dipoles. In this letter, we propose a new dipole source reconstruction method that leverages the neural network (NN) framework to address this challenge effectively. Unlike most NN applications focused on regression or classification tasks, we employ the NN as an optimizer. We incorporate a square function into the NN model to effectively handle the phaseless near-field. The input to the NN is the Green's function of the dipole array, while the output is the squared magnitude of the predicted near-field. The parameters in the NN represent the dipole moments to be optimized. This method provides a new pathway for dipole source reconstruction in phaseless near-field scanning.