Electromagnetic Green’s Function Retrieval of 3D Environment Using Physical-driven Neural Point Field

Electromagnetic (EM) Green’s function retrieval is directly related to the representation of wireless channel response, which is of fundamental importance in diverse areas including communication, sensing, navigation, etc., but remains an open challenge especially in dealing with the intricate large-scale environment due to the inherent complex wave-matter interaction. Since the recent breakthrough of neural radiance field (NeRF) models, many data-driven approaches have been developed to address the above formidable challenge, however, they have to struggle with the trade-off among accuracy, efficiency, interpretability, and scalability. In order to resolve this issue, we here report a physical-driven neural point field representation of Green’s function: i) to represent the 3D scene with a collection of surrogate sources, ii) to approximate the Green’s function by characterizing every virtual source with a simple but physically meaningful Green’s function along with an attention network, and iii) to train the parameters involved from the sparse measurements available. Selected results have been provided to demonstrate that, in comparison with existing NeRF-based models, our method has concise physical interpretability, and is capable of remarkably reducing the training time while maintaining the accuracy, and thus provides us with a scalable efficient solution to the intricate large-scale environments. We faithfully expect that our strategy could find its valuable applications in the emerging systems of integrated sensing and communication. Our code for reproducing the results in this work is available at https://github.com/Phoenix0qi/GreenRetrival.git.