Neural Reflectance Fields for Radio-Frequency Ray Tracing

TL;DR

The paper tackles scalable RF material reflectivity estimation for ray tracing by introducing neural reflectance fields in the RF domain, implemented as a complex-valued MLP $F_{ extbf{w}}:(\theta,s)\rightarrow (\Delta A(\theta,s), \Delta \Theta(\theta,s))$ with $\delta(\theta,s)=\Delta A(\theta,s)e^{j\Delta \Theta(\theta,s)}$. A differentiable RF ray-tracing framework integrates traditional ray tracing with the neural reflectance field to predict per-ray and per-transmitter receive powers, using the rendering relation $P^{rx}_d = \left| \sum_{i} \sqrt{P^{tx}_{i}} \sum_{k\in\mathcal{R}_{i,d}} H^{FS}_k \delta_k \right|^2$ where $H^{FS}_k=\frac{\lambda_c}{4\pi d_k} e^{j2\pi d_k/\lambda_c}$ and $\delta_k=\prod \Delta A(\theta, s) e^{j\Delta \Theta(\theta,s)}$. Experiments on a campus-scale OpenStreetMap scene show that the approach outperforms NeRF$^2$ in total receive-power prediction, achieves accurate ray-level attenuation estimates, and enables reliable SINR and interference-channel prediction under varying data densities, highlighting its potential for scalable RF propagation modeling and resource allocation tasks.

Abstract

Ray tracing is widely employed to model the propagation of radio-frequency (RF) signal in complex environment. The modelling performance greatly depends on how accurately the target scene can be depicted, including the scene geometry and surface material properties. The advances in computer vision and LiDAR make scene geometry estimation increasingly accurate, but there still lacks scalable and efficient approaches to estimate the material reflectivity in real-world environment. In this work, we tackle this problem by learning the material reflectivity efficiently from the path loss of the RF signal from the transmitters to receivers. Specifically, we want the learned material reflection coefficients to minimize the gap between the predicted and measured powers of the receivers. We achieve this by translating the neural reflectance field from optics to RF domain by modelling both the amplitude and phase of RF signals to account for the multipath effects. We further propose a differentiable RF ray tracing framework that optimizes the neural reflectance field to match the signal strength measurements. We simulate a complex real-world environment for experiments and our simulation results show that the neural reflectance field can successfully learn the reflection coefficients for all incident angles. As a result, our approach achieves better accuracy in predicting the powers of receivers with significantly less training data compared to existing approaches.

Figures (6)

• Figure 1: Our proposed RF ray tracing framework with the neural reflectance field. Our framework includes three components: (1) Ray tracing module, which employs current ray tracing algorithms to discover the strong rays for each TX-RX pair, maps the reflection points to their surface IDs, and steams the tuples, (incident angle, surface ID), to the neural reflectance fields. (2) Neural reflectance field, which takes as input the tuples and outputs the attenuation and phase change of the RF signal. (3) Rendering module, which composes point-level attenuation into ray-level attenuation, aggregates across multiple rays to compute the received signal for each TX-RX pair, and sums up the received signal from all transmitters to compute the receive power.
• Figure 2: Learning the relation between reflection attenuation and the angle of incidence with the neural reflectance field. Incident rays may differ in density at different angles, resulting in sparse regions to have larger learning errors for the reflection attenuation.
• Figure 3: Campus-level 3D scene geometry from OpenStreetMap
• Figure 4: Accuracy of the neural reflectance field
• Figure 5: Comparison with existing approaches