NeRF2: Neural Radio-Frequency Radiance Fields

Abstract

Although Maxwell discovered the physical laws of electromagnetic waves 160\nyears ago, how to precisely model the propagation of an RF signal in an\nelectrically large and complex environment remains a long-standing problem. The\ndifficulty is in the complex interactions between the RF signal and the\nobstacles (e.g., reflection, diffraction, etc.). Inspired by the great success\nof using a neural network to describe the optical field in computer vision, we\npropose a neural radio-frequency radiance field, NeRF$^\\textbf{2}$, which\nrepresents a continuous volumetric scene function that makes sense of an RF\nsignal's propagation. Particularly, after training with a few signal\nmeasurements, NeRF$^\\textbf{2}$ can tell how/what signal is received at any\nposition when it knows the position of a transmitter. As a physical-layer\nneural network, NeRF$^\\textbf{2}$ can take advantage of the learned statistic\nmodel plus the physical model of ray tracing to generate a synthetic dataset\nthat meets the training demands of application-layer artificial neural networks\n(ANNs). Thus, we can boost the performance of ANNs by the proposed\nturbo-learning, which mixes the true and synthetic datasets to intensify the\ntraining. Our experiment results show that turbo-learning can enhance\nperformance with an approximate 50% increase. We also demonstrate the power of\nNeRF$^\\textbf{2}$ in the field of indoor localization and 5G MIMO.\n