Bidirectional Neural Radio-Frequency Radiance Fields for 6-D Channel Knowledge Map Construction

Abstract

This letter studies the construction of six-dimensional (6D) any-to-any (X2X) channel state information (CSI) maps, a specific type of channel knowledge map (CKM), for single-input multiple-output (SIMO) systems with varying transmitter (Tx) and receiver (Rx) locations, using limited measurements. We propose a novel bidirectional neural radio-frequency radiance field (BiNeRF2) method. Specifically, the SIMO channel is characterized based on a last-bounce propagation model, where an explicit physical model is employed to describe the last-bounce propagation, while an implicit BiNeRF2 model is used to capture the preceding propagation toward the last-interaction point. In the BiNeRF2 model, propagation-related parameters are learned by a neural network from spatial and bidirectional angular geometric information. Experiments in indoor environments demonstrate that BiNeRF2 outperforms neural radio-frequency radiance fields (NeRF2) and wireless radiation field Gaussian splatting (WRF-GS) in 3D settings, while providing competitive 6D prediction accuracy and low inference latency.