Neural Gaussian Radio Fields for Channel Estimation

TL;DR

This work tackles the CSI estimation bottleneck by introducing neural Gaussian radio fields (nGRF), a physics-informed neural field that represents the radio environment as a sum of learnable 3D Gaussian primitives with complex-valued responses. By replacing occlusion-based rendering with direct complex aggregation that enforces wave superposition, nGRF reframes channel estimation as a structured inverse problem with strong physical priors, enabling millisecond inference and significantly reduced pilot overhead. Across indoor and outdoor scenarios, including field measurements, nGRF delivers substantial SNR gains (up to >28 dB in some cases) and outperforms NeRF-based and data-driven baselines while requiring far less data and training time. The approach generalizes to other linear wave domains by substituting the Green's kernel, offering a scalable blueprint for physics-informed neural fields in acoustics, seismology, and beyond.

Abstract

Accurate channel state information (CSI) is a critical bottleneck in modern wireless networks, with pilot overhead consuming 11\% to 21\% of transmission bandwidth and feedback delays causing severe throughput degradation under mobility. Addressing this requires rethinking how neural fields represent coherent wave phenomena. This work introduces \textit{neural Gaussian radio fields (\textcolor{stanfordred}{nGRF})}, a physics-informed framework that fundamentally reframes neural field design by replacing view-dependent rasterization with direct complex-valued aggregation in 3D space. This approach natively models wave superposition rather than visual occlusion. The architectural shift transforms the learning objective from function-fitting to source-recovery, a well-posed inverse problem grounded in electromagnetic theory. While demonstrated for wireless channel estimation, the core principle of explicit primitive-based fields with physics-constrained aggregation extends naturally to any coherent wave-based domain, including acoustic propagation, seismic imaging, and ultrasound reconstruction. Evaluations show that the inductive bias of \textcolor{stanfordred}{nGRF} achieves 10.9 dB higher prediction SNR than state-of-the-art methods with 220$\times$ faster inference (1.1 ms vs. 242 ms), 18$\times$ lower measurement density, and 180$\times$ faster training. For large-scale outdoor environments where implicit methods fail, \textcolor{stanfordred}{nGRF} achieves 28.32 dB SNR, demonstrating that structured representations supplemented by domain physics can fundamentally outperform generic deep learning architectures.

Figures (4)

• Figure 1: Neural architecture. The AttributeNetwork processes Gaussian and transmitter positions through separate positional encoders to produce latent features and base activation logits. The DecoderNetwork then transforms the latent features into complex channel contributions.
• Figure 2: Channel magnitude response. Comparison between predicted (left) and reference PACE estimate (right) channel magnitude response across subcarriers and transmit antennas in the outdoor environment using nGRF.
• Figure 3: Multipath propagation with path-dependent attenuation and phase.
• Figure 4: 3D environments used for ray tracing.