Neural Gaussian Radio Fields (nGRF)

• nGRF are explicit neural representations that model electromagnetic channels using anisotropic 3D Gaussian primitives to capture localized RF behaviors.
• They employ physics-driven forward operators, including differentiable ray tracing and BRDF-like scattering, to aggregate complex channel contributions.
• nGRF approaches significantly reduce pilot overhead and inference latency while enabling multimodal reconstruction of both radio and optical scene properties.

Neural Gaussian Radio Fields (nGRF) are neural wireless-field representations built from explicit Gaussian primitives rather than a single implicit volumetric MLP. In the most direct formulation, an nGRF models the wireless environment as a finite set of anisotropic 3D Gaussian “radio primitives,” each acting as a localized radio modulator, and synthesizes the channel between a transmitter and a receiver by directly aggregating the contributions of those Gaussians in 3D space (Umer et al., 6 Aug 2025). Closely related formulations use Gaussian primitives to reconstruct wireless radiation fields, spatial spectra, radio maps, or physically grounded inverse-rendered RF scenes, often sharing geometry with optical radiance fields and using differentiable rendering or ray tracing to map scene structure to received signals (Wen et al., 2024, Wang et al., 8 Apr 2026, Wen et al., 27 Jan 2026).

1. Definition and conceptual scope

The term “Neural Gaussian Radio Fields” is most explicitly defined in work on channel estimation, where the scene is represented by $N$ anisotropic 3D Gaussian primitives $\{G_i\}_{i=1}^N$, and the MIMO channel matrix is rendered by a direct 3D superposition: $$\mathbf{H}(\mathbf{p}_{\mathrm{rx}}, \mathbf{p}_{\mathrm{tx}}) = \sum_{i=1}^N w_i(\mathbf{p}_{\mathrm{rx}}, \boldsymbol{\mu}_i, \boldsymbol{\Sigma}_i)\, \mathbf{C}_i,$$ with $w_i$ a Gaussian spatial weight and $\mathbf{C}_i \in \mathbb{C}^{N_t \times N_r}$ the complex channel contribution of primitive $i$ (Umer et al., 6 Aug 2025). In that formulation, nGRF is an explicit 3D-geometry-aware neural representation of the radio channel, intended to replace dense pilot-based CSI acquisition with a compact learned model of the environment’s electromagnetic field (Umer et al., 6 Aug 2025).

The broader literature uses related but not always identical terminology. NeRF$^2$ introduced a continuous neural radio-frequency radiance field in which a NeRF-like MLP maps transmitter position, voxel position, and direction to complex attenuation and complex retransmitted field, then integrates along rays to obtain received RF observations (Zhao et al., 2023). Subsequent Gaussian-based formulations replaced ray-marched implicit fields with explicit Gaussian primitives: WRF-GS framed wireless radiation field reconstruction with 3D Gaussian splatting (Wen et al., 2024), RF-3DGS treated Gaussians as carriers of radio spatial spectra and Spatial-CSI (Zhang et al., 2024), SwiftWRF moved the Gaussian basis into the angular domain with deformable 2D Gaussians (Liu et al., 15 Jun 2025), URF-GS unified optical and wireless rendering in one Gaussian scene representation (Wen et al., 27 Jan 2026), and RFIR proposed a physically grounded RF inverse-rendering blueprint in which Gaussian primitives encode geometry and electromagnetic response while RF emission is supplied explicitly by the forward model (Wang et al., 8 Apr 2026).

A central conceptual distinction separates nGRF-like models from pure NeRF-style wireless radiance fields. In NeRF$^2$, the field is an implicit scene function learned per scene (Zhao et al., 2023). In nGRF-style systems, the field is explicit: Gaussian primitives carry geometric, radiometric, or electromagnetic attributes, and the forward operator is expressed as Gaussian aggregation, Gaussian splatting, BRDF-/BSDF-like scattering, or differentiable ray tracing (Umer et al., 6 Aug 2025, Wang et al., 8 Apr 2026).

2. Gaussian primitives and field parameterizations

The simplest nGRF parameterization uses anisotropic 3D Gaussians with centers $\boldsymbol{\mu}_i \in \mathbb{R}^3$, covariance matrices $\boldsymbol{\Sigma}_i \in \mathbb{R}^{3\times 3}$, and spatial weights

$\{G_i\}_{i=1}^N$0

where $\{G_i\}_{i=1}^N$1 is a learned base activation (Umer et al., 6 Aug 2025). Covariances are commonly parameterized as

$\{G_i\}_{i=1}^N$2

with $\{G_i\}_{i=1}^N$3 obtained from a unit quaternion and $\{G_i\}_{i=1}^N$4 diagonal with positive scales (Umer et al., 6 Aug 2025). This makes the primitive an explicit anisotropic basis element in physical 3D space.

Physically grounded inverse-rendering variants enrich each primitive with material and surface attributes. In RFIR, a Gaussian primitive is

$\{G_i\}_{i=1}^N$5

where $\{G_i\}_{i=1}^N$6 is the local surface normal, $\{G_i\}_{i=1}^N$7 modulates visibility and occlusion, $\{G_i\}_{i=1}^N$8 is an effective roughness parameter, and $\{G_i\}_{i=1}^N$9 is a complex reflection coefficient with learned magnitude and phase (Wang et al., 8 Apr 2026). That decomposition explicitly separates geometry $$\mathbf{H}(\mathbf{p}_{\mathrm{rx}}, \mathbf{p}_{\mathrm{tx}}) = \sum_{i=1}^N w_i(\mathbf{p}_{\mathrm{rx}}, \boldsymbol{\mu}_i, \boldsymbol{\Sigma}_i)\, \mathbf{C}_i,$$0, material electromagnetic behavior $$\mathbf{H}(\mathbf{p}_{\mathrm{rx}}, \mathbf{p}_{\mathrm{tx}}) = \sum_{i=1}^N w_i(\mathbf{p}_{\mathrm{rx}}, \boldsymbol{\mu}_i, \boldsymbol{\Sigma}_i)\, \mathbf{C}_i,$$1, and occlusion $$\mathbf{H}(\mathbf{p}_{\mathrm{rx}}, \mathbf{p}_{\mathrm{tx}}) = \sum_{i=1}^N w_i(\mathbf{p}_{\mathrm{rx}}, \boldsymbol{\mu}_i, \boldsymbol{\Sigma}_i)\, \mathbf{C}_i,$$2 (Wang et al., 8 Apr 2026).

URF-GS uses a different but related attribute set. Each 3D Gaussian primitive

$$\mathbf{H}(\mathbf{p}_{\mathrm{rx}}, \mathbf{p}_{\mathrm{tx}}) = \sum_{i=1}^N w_i(\mathbf{p}_{\mathrm{rx}}, \boldsymbol{\mu}_i, \boldsymbol{\Sigma}_i)\, \mathbf{C}_i,$$3

combines geometry and optical attributes with radio-related material attributes: scattering albedo $$\mathbf{H}(\mathbf{p}_{\mathrm{rx}}, \mathbf{p}_{\mathrm{tx}}) = \sum_{i=1}^N w_i(\mathbf{p}_{\mathrm{rx}}, \boldsymbol{\mu}_i, \boldsymbol{\Sigma}_i)\, \mathbf{C}_i,$$4, metallic $$\mathbf{H}(\mathbf{p}_{\mathrm{rx}}, \mathbf{p}_{\mathrm{tx}}) = \sum_{i=1}^N w_i(\mathbf{p}_{\mathrm{rx}}, \boldsymbol{\mu}_i, \boldsymbol{\Sigma}_i)\, \mathbf{C}_i,$$5, roughness $$\mathbf{H}(\mathbf{p}_{\mathrm{rx}}, \mathbf{p}_{\mathrm{tx}}) = \sum_{i=1}^N w_i(\mathbf{p}_{\mathrm{rx}}, \boldsymbol{\mu}_i, \boldsymbol{\Sigma}_i)\, \mathbf{C}_i,$$6, and a per-Gaussian normal vector $$\mathbf{H}(\mathbf{p}_{\mathrm{rx}}, \mathbf{p}_{\mathrm{tx}}) = \sum_{i=1}^N w_i(\mathbf{p}_{\mathrm{rx}}, \boldsymbol{\mu}_i, \boldsymbol{\Sigma}_i)\, \mathbf{C}_i,$$7 (Wen et al., 27 Jan 2026). The same Gaussian set supports both RGB rendering and radio propagation, making the scene representation shared across modalities (Wen et al., 27 Jan 2026).

Not all nGRF-like systems place Gaussians in 3D Euclidean space. SwiftWRF uses deformable 2D Gaussian primitives in the angular plane $$\mathbf{H}(\mathbf{p}_{\mathrm{rx}}, \mathbf{p}_{\mathrm{tx}}) = \sum_{i=1}^N w_i(\mathbf{p}_{\mathrm{rx}}, \boldsymbol{\mu}_i, \boldsymbol{\Sigma}_i)\, \mathbf{C}_i,$$8, with center $$\mathbf{H}(\mathbf{p}_{\mathrm{rx}}, \mathbf{p}_{\mathrm{tx}}) = \sum_{i=1}^N w_i(\mathbf{p}_{\mathrm{rx}}, \boldsymbol{\mu}_i, \boldsymbol{\Sigma}_i)\, \mathbf{C}_i,$$9, covariance $w_i$0, attenuation $w_i$1, and a two-channel signal vector $w_i$2 encoding real and imaginary parts of the local complex response (Liu et al., 15 Jun 2025). In that case the wireless field is represented directly in the receiver’s angular observation domain rather than in scene coordinates.

3. Forward operators and physical modeling

A major fault line within the nGRF literature concerns the rendering equation itself. Direct channel-field aggregation models avoid optical-style visibility compositing altogether. In the channel-estimation nGRF formulation, the field obeys direct 3D electromagnetic aggregation: the channel matrix is a sum of Gaussian-weighted complex matrices, and the method explicitly argues that radio does not behave like alpha-composited optical appearance and that electromagnetic fields superpose in 3D rather than on an image plane (Umer et al., 6 Aug 2025). That critique is directed at adapting standard 3D Gaussian splatting too literally to RF.

RFIR takes the opposite route: it retains Gaussian primitives but replaces learned “radiance” with a physically grounded RF rendering equation. The outgoing complex field of a Gaussian is expressed as a scattering integral over sampled incident directions,

$w_i$3

with an RF-aware BSDF

$w_i$4

and Friis-like incident-field modeling from transmitter to Gaussian, followed by alpha-blended accumulation from Gaussians to the receiver (Wang et al., 8 Apr 2026). In that formulation, the current implementation is a single-bounce model with explicit free-space path loss, phase accumulation, BSDF-based scattering, and volumetric occlusion (Wang et al., 8 Apr 2026).

GaussianRT pushes further toward deterministic propagation. It embeds planar Gaussian primitives into a hardware-accelerated ray tracing structure, performs coarse path search, Fermat-principle refinement, and path pruning, then computes RF transfer functions, channel frequency responses, CIRs, PDPs, RMS delay spreads, and RSS from the resulting paths (Vaara et al., 8 May 2026). Path amplitudes incorporate Fresnel reflection coefficients and antenna patterns, and the CFR is assembled as

$w_i$5

with $w_i$6 determined by Gaussian-defined path lengths (Vaara et al., 8 May 2026). This line treats Gaussian scenes as ray-intersectable geometry rather than merely as splatted blobs.

URF-GS occupies an intermediate position. It reuses the same Gaussian geometry for both optics and radio, but the RF forward model is based on a PBR-like BRDF decomposition $w_i$7, free-space path loss over path segments, and a differentiable path tracer operating on a Gaussian-derived G-buffer (Wen et al., 27 Jan 2026). A plausible implication is that nGRF research is converging on two distinct but compatible views: Gaussian primitives as explicit field bases, and Gaussian primitives as differentiable geometric supports for electromagnetic transport.

4. Optimization regimes and representative systems

Most nGRF-like systems are trained by inverse rendering, but the supervision modality and optimization schedule vary sharply across models.

System	Gaussian domain	Main supervised output
WRF-GS / WRF-GS+ (Wen et al., 2024)	3D scene Gaussians	Spatial spectrum, RSSI, CSI
RF-3DGS (Zhang et al., 2024)	3D scene Gaussians	AoA spectrum and Spatial-CSI
SwiftWRF (Liu et al., 15 Jun 2025)	2D angular Gaussians	Complex spatial spectrum
nGRF channel field (Umer et al., 6 Aug 2025)	3D scene Gaussians	Complex MIMO channel matrix
URF-GS (Wen et al., 27 Jan 2026)	Shared 3D radio-optical Gaussians	Spatial spectrum / 3D radio map
RFIR (Wang et al., 8 Apr 2026)	3D RF Gaussian primitives	RCS, RSSI, scene editability

A common training pattern is two-stage optimization. RFIR first reconstructs geometry from multi-view RGB with standard 3DGS plus geometric regularization, then freezes geometry and learns RF attributes $w_i$8 from RF measurements using

$w_i$9

as its RF inverse-rendering loss (Wang et al., 8 Apr 2026). URF-GS follows a nearly identical pattern: an optical stage learns geometry, normals, and color under RGB, depth, and normal losses, then a wireless stage freezes geometry and optimizes material attributes $\mathbf{C}_i \in \mathbb{C}^{N_t \times N_r}$0 to match spatial spectrum images (Wen et al., 27 Jan 2026).

RF-3DGS likewise uses a two-stage fusion pipeline. Visual data train the Gaussian geometry backbone, after which geometry is frozen and density plus CSI-encoded spherical-harmonic coefficients are optimized with radio supervision (Zhang et al., 2024). SwiftWRF differs by training a canonical 2D Gaussian set first and then learning a deformation MLP that predicts residual center, signal, and attenuation updates as a function of transceiver position (Liu et al., 15 Jun 2025).

A related Gaussian RF line avoids learned field rendering almost entirely. “Active Sampling and Gaussian Reconstruction for Radio Frequency Radiance Field” models virtual RF sources with an RBF covariance, performs Gaussian conditioning to obtain posterior mean and variance at target points, and uses posterior variance for active sampling and quasi-dynamic difference-field updates (Gau et al., 2024). Although that work is training-free rather than a neural Gaussian field in the strict sense, it supplies an uncertainty-aware counterpart to deterministic nGRF formulations.

5. Empirical performance and application domains

The most aggressive empirical claims come from the channel-estimation nGRF formulation. In indoor scenarios it achieves a $\mathbf{C}_i \in \mathbb{C}^{N_t \times N_r}$1 higher prediction SNR than state-of-the-art methods, reduces inference latency from 242 ms to 1.1 ms, requires only 0.011 measurements per cubic foot compared to 0.2–178.1 for existing methods, and reduces training time from hours to minutes; in large-scale outdoor environments it reports an SNR of 26.2 dB (Umer et al., 6 Aug 2025). Inference and training speedups are attributed to explicit Gaussian aggregation rather than volumetric ray marching or 2D projection-based RF splatting (Umer et al., 6 Aug 2025).

At the system level, 6G Twin extends Gaussian Radio Fields into a RAN stack that couples compressed CSI acquisition, continual channel prediction, and nonlinear precoding. In that framework, GRF replaces dense pilots with a sparse Gaussian field, cutting pilot overhead by about $\mathbf{C}_i \in \mathbb{C}^{N_t \times N_r}$2 while delivering 1.1 ms inference and less than 2 minutes on-site training; a replay-driven continual learner improves channel NMSE by more than 10 dB over frozen predictors and an additional 2–5 dB over uniform replay (Mohsin et al., 23 Sep 2025). This moves nGRF from a channel-modeling primitive toward an operational CSI substrate.

Physically grounded inverse-rendering variants emphasize different tasks. RFIR generalizes radar cross-section synthesis, RSSI prediction, and wireless scene editability; it reports RCS MAE in the range 1.87–2.69 dB, wideband RCS MAE of 1.69 dB across a 160 MHz band, classroom RSSI MAE of 4.82 dB, and spatial RCS extrapolation to 3 m with MAE 2.62 dB (Wang et al., 8 Apr 2026). The same work reports that removing explicit path loss and phase substantially worsens RCS error, and removing learned LoS/NLoS mixing reduces RSSI accuracy by about 35%, underscoring the importance of explicit propagation structure (Wang et al., 8 Apr 2026).

Unified multimodal fields show a different strength profile. URF-GS reports up to a 24.7% improvement in spatial spectrum prediction accuracy and a $\mathbf{C}_i \in \mathbb{C}^{N_t \times N_r}$3 increase in sample efficiency for 3D radio map construction compared with NeRF-based methods (Wen et al., 27 Jan 2026). In full-training, few-shot, and zero-shot transmitter generalization settings, it maintains the highest PSNR among the listed baselines for spatial spectrum prediction, including 17.38 in the full setting and 9.33 in the zero-shot setting (Wen et al., 27 Jan 2026). This suggests that shared optical-radio geometry can regularize wireless prediction when RF supervision is sparse.

RF-3DGS focuses on radio radiance field reconstruction with explicit Gaussian splatting. It reports about 3 minutes of training, about 2 ms per query spectrum, and an 84.64% LPIPS reduction relative to NeRF$\mathbf{C}_i \in \mathbb{C}^{N_t \times N_r}$4 on reconstructed AoA spectra, while also producing fine-grained Spatial-CSI including channel gain, delay, AoA, and AoD (Zhang et al., 2024). SwiftWRF pushes angular-domain Gaussian fields toward real-time operation, reporting over 100000 fps spectrum rendering, up to $\mathbf{C}_i \in \mathbb{C}^{N_t \times N_r}$5 speedup over existing state-of-the-art methods, and improved AoA and RSSI prediction; in an RSSI benchmark it reports median L1 error of 2.91 dB versus 3.05 dB for NeRF$\mathbf{C}_i \in \mathbb{C}^{N_t \times N_r}$6 and 6.74 dB for MRI (Liu et al., 15 Jun 2025). GaussianRT, finally, shows that visually reconstructed Gaussian scenes can support path-level RF simulation: in a corridor scene it reports RMS delay spread 1.56 ns versus 1.67 ns in measurements, and in a smartphone-captured lab scene it reports RSS MAE of about 0.2–0.7 dB on held-out receivers after RF calibration (Vaara et al., 8 May 2026).

6. Limitations, misconceptions, and research directions

A recurring misconception is that nGRF is simply standard Gaussian splatting applied to radio. The literature argues otherwise. The direct-aggregation channel-estimation model explicitly rejects non-physical 2D projection and alpha-blending as a general RF rendering principle, emphasizing that electromagnetic fields superpose in 3D and that visibility in the optical sense is not the right primitive for channel synthesis (Umer et al., 6 Aug 2025). By contrast, RFIR, URF-GS, and GaussianRT retain projection, compositing, or path tracing, but they modify those operators with Friis-like propagation, BRDF/BSDF terms, Fresnel coefficients, or deterministic path refinement (Wang et al., 8 Apr 2026, Vaara et al., 8 May 2026).

A second misconception is that Gaussian primitives must correspond one-to-one with physical scatterers. That is model-dependent. The channel-estimation nGRF formulation reports that LiDAR-geometry initialization hurts SNR because the Gaussians are field bases, not literal scatterers, and because constraining them to surfaces reduces their flexibility as an electromagnetic basis (Umer et al., 6 Aug 2025). Conversely, RFIR and GaussianRT deliberately endow Gaussians with surface normals, roughness, or semantic/material labels so that they can function as local electromagnetic patches or ray-intersectable surfaces (Wang et al., 8 Apr 2026, Vaara et al., 8 May 2026).

The dominant limitations are consistent across papers. Many models assume static scenes (Umer et al., 6 Aug 2025, Wen et al., 27 Jan 2026). RFIR does not explicitly model polarization, treats $\mathbf{C}_i \in \mathbb{C}^{N_t \times N_r}$7 as a scalar reflection coefficient, and uses single-bounce scattering as its main mechanism (Wang et al., 8 Apr 2026). URF-GS uses a simplified EM model based on PBR-style BRDF plus FSPL rather than full Maxwell equations and does not explicitly model diffraction or polarization (Wen et al., 27 Jan 2026). The channel-estimation nGRF notes hyperparameter sensitivity, lack of explicit time dynamics, environment-specific training, and challenges under very large outdoor scenes or severe noise (Umer et al., 6 Aug 2025).

The main research directions are already visible. Dynamic nGRF variants are repeatedly tied to 4D Gaussian splatting (Wen et al., 27 Jan 2026). RFIR proposes frequency-conditioned deformation of roughness and reflection parameters for wideband modeling and suggests extensions to polarization, multi-bounce, and diffraction kernels (Wang et al., 8 Apr 2026). 6G Twin points toward uncertainty-aware pilot placement, robust scheduling with predictor covariances, wideband and near-field support, and multi-cell cooperative Gaussian twins (Mohsin et al., 23 Sep 2025). A plausible implication is that the field is moving from scene-specific RF rendering toward general-purpose wireless world models in which geometry, materials, spectrum, mobility, and control are all mediated through Gaussian primitives.