VoxelRF: Voxelized Radiance Field for Fast Wireless Channel Modeling

Wireless channel modeling in complex environments is crucial for wireless communication system design and deployment. Traditional channel modeling approaches face challenges in balancing accuracy, efficiency, and scalability, while recent neural approaches such as neural radiance field (NeRF) suffer from long training and slow inference. To tackle these challenges, we propose voxelized radiance field (VoxelRF), a novel neural representation for wireless channel modeling that enables fast and accurate synthesis of spatial spectra. VoxelRF replaces the costly multilayer perception (MLP) used in NeRF-based methods with trilinear interpolation of voxel grid-based representation, and two shallow MLPs to model both propagation and transmitter-dependent effects. To further accelerate training and improve generalization, we introduce progressive learning, empty space skipping, and an additional background entropy loss function. Experimental results demonstrate that VoxelRF achieves competitive accuracy with significantly reduced computation and limited training data, making it more practical for real-time and resource-constrained wireless applications.

💡 Research Summary

The paper introduces VoxelRF, a voxel‑based neural representation for wireless channel modeling that dramatically speeds up both training and inference while preserving high‑fidelity spatial spectra. Traditional channel models (deterministic ray‑tracing, stochastic, or hybrid) either require exhaustive environmental knowledge or sacrifice accuracy. Recent neural approaches such as NeRF‑2 adapt the neural radiance field concept to wireless propagation, but they inherit NeRF’s reliance on a deep multilayer perceptron (MLP) for every sample point, leading to long training times and slow inference.

VoxelRF replaces the deep MLP with two explicit 3‑D voxel grids: an attenuation grid V_attn and a feature grid V_feat. For any sampled point x_i along a ray, trilinear interpolation yields the local attenuation σ(x_i) and a feature vector Feat(x_i). Because interpolation is differentiable, gradients flow directly to the voxel values, allowing end‑to‑end learning of the volumetric field. The voxel representation is naturally sparse—most of the volume is air—so VoxelRF introduces an “empty‑space skipping” mechanism: if the interpolated attenuation falls below a threshold τ (1e‑4), the sample is discarded, avoiding unnecessary MLP evaluations.

To handle the dynamic nature of a moving transmitter (Tx), a shallow deformation network D predicts a position‑dependent offset ΔFeat(x_i). The final feature fed to the radiance network I is Feat(x_i)+ΔFeat(x_i), effectively decoupling static material properties (stored in the voxels) from Tx‑induced multipath variations. The radiance network I is also shallow (two layers, 256 hidden units) and outputs the emitted signal S(x_i,−ω) using a sigmoid activation, while σ(x_i) uses a softplus to keep physical meaning. Positional encoding is applied to Tx position, sample position, and view direction (L=5 for positions, L=4 for directions) to capture high‑frequency variations.

Training employs a progressive‑learning schedule: the voxel grid starts at half the final resolution and is up‑sampled by a factor of two over M stages, reducing over‑fitting risk and stabilizing convergence. The loss combines a mean‑squared error on the per‑direction spatial spectrum (L_spectrum) with a background entropy loss (L_bg) that encourages confident predictions of occupied versus empty space. The total loss is L = L_spectrum + λ_bg·L_bg with λ_bg = 1e‑4.

Experiments use a real‑world RFID dataset (915 MHz, 4×4 antenna array, 6 123 Tx positions) in a 23 m × 24 m × 2 m indoor lab. VoxelRF is configured with 160³ voxels, feature dimension 24, and trained on a single RTX 3080 Ti GPU for ~100 k iterations (≈1 hour). Baselines include NeRF‑2, WRF‑GS+ (Gaussian‑splatting based), and a DCGAN generator. VoxelRF consistently achieves higher structural similarity index (SSIM) scores across four test Tx locations, preserving sharp transitions and high‑frequency fading patterns that the Gaussian primitives of WRF‑GS+ blur. Even when training data is reduced to 10 % of the full set, VoxelRF maintains median SSIM above 0.85, demonstrating strong data efficiency.

In summary, VoxelRF shows that a voxel‑grid combined with shallow MLPs, empty‑space skipping, progressive resolution growth, and an entropy‑based regularizer can deliver fast, accurate, and data‑efficient wireless channel synthesis. This makes real‑time channel prediction feasible on resource‑constrained platforms and opens the door for extensions to multi‑frequency, multi‑Tx, and mobile edge scenarios in upcoming 6G and beyond networks.