Neural Gaussian Radio Fields for Channel Estimation

Accurate channel state information (CSI) remains the most critical bottleneck in modern wireless networks, with pilot overhead consuming up to 11-21% of transmission bandwidth, increasing latency by 20-40% in massive MIMO systems, and reducing potential spectral efficiency by over 53%. Traditional estimation techniques fundamentally fail under mobility, with feedback delays as small as 4 ms causing 50% throughput degradation at even modest speeds (30 km/h). We present neural Gaussian radio fields (nGRF), a novel framework that leverages explicit 3D Gaussian primitives to synthesize complex channel matrices accurately and efficiently. Unlike NeRF-based approaches that rely on slow implicit representations or existing Gaussian splatting methods that use non-physical 2D projections, nGRF performs direct 3D electromagnetic field aggregation, with each Gaussian acting as a localized radio modulator. nGRF demonstrates superior performance across diverse environments: in indoor scenarios, it achieves a 10.9$\times$ higher prediction SNR than state of the art methods while reducing inference latency from 242 ms to just 1.1 ms (a 220$\times$ speedup). For large-scale outdoor environments, where existing approaches fail to function, nGRF achieves an SNR of 26.2 dB. Moreover, nGRF requires only 0.011 measurements per cubic foot compared to 0.2-178.1 for existing methods, thereby reducing data collection burden by 18$\times$. Training time is similarly reduced from hours to minutes (a 180$\times$ reduction), enabling rapid adaptation to dynamic environments. The code and datasets are available at: https://github.com/anonym-auth/n-grf