Source-linked AI summary
NeWRF: A Deep Learning Framework for Wireless Radiation Field Reconstruction and Channel Prediction
Haofan Lu, Christopher Vattheuer, Baharan Mirzasoleiman, Omid Abari
TL;DR
Sparse site surveys are costly and leave wireless channel quality unknown at unvisited locations. NeWRF adapts NeRF to learn wireless radiation scenes from sparse measurements, and it outperforms baselines across three environments.
Problem
Dense site surveys are time-consuming and costly, while sparse random measurements leave channel quality at unvisited locations unknown and can miss dead spots.
Source: §1 · p. 1Method
NeWRF adapts NeRF to wireless propagation, using sparse channel measurements and direction-of-arrival-guided ray casting to predict channels at unvisited locations.
Source: §1 · p. 2Results
NeWRF outperforms all baseline methods in all three environments, and inspection of its learned model reveals scene simplicity that supports sparse-measurement learning in complex room-scale spaces.
Source: §1 · p. 2, §5.1 · p. 8Takeaways & Limitations
NeWRF is a first step toward practical NeRF use in wireless applications, with potential opportunities to optimize future wireless-network performance.
Source: §7 · p. 9The narrowband-channel approach does not perform well when phase measurements are inaccurate or unavailable.
Source: §6 · p. 9
Abstract
from arXiv · showhide
We present NeWRF, a deep learning framework for predicting wireless channels. Wireless channel prediction is a long-standing problem in the wireless community and is a key technology for improving the coverage of wireless network deployments. Today, a wireless deployment is evaluated by a site survey which is a cumbersome process requiring an experienced engineer to perform extensive channel measurements. To reduce the cost of site surveys, we develop NeWRF, which is based on recent advances in Neural Radiance Fields (NeRF). NeWRF trains a neural network model with a sparse set of channel measurements, and predicts the wireless channel accurately at any location in the site. We introduce a series of techniques that integrate wireless propagation properties into the NeRF framework to account for the fundamental differences between the behavior of light and wireless signals. We conduct extensive evaluations of our framework and show that our approach can accurately predict channels at unvisited locations with significantly lower measurement density than prior state-of-the-art
1. Introduction
Dense site surveys are costly, while sparse measurements leave signal quality at unvisited locations unknown. NeWRF adapts NeRF to reconstruct wireless channels from sparse measurements while addressing differences between light and wireless propagation.
1. Introduction
- Exhaustive surveys require dense measurements and are rarely performed because of their time and cost; sparse random surveys leave unvisited signal quality—and potential dead spots—unknown.Effective surveys require an experienced engineer to measure a dense grid.Source: §1 · p. 1
- Fine-scale channel variation from constructive and destructive interference makes location-to-channel learning without propagation physics dependent on dense measurements.Signals arriving by different paths interfere, producing spatial variation.Source: Figure 1 · p. 1, §1 · p. 2
- Adapting NeRF to wireless requires handling room-scale scenes, low-information complex-valued measurements, all-direction antenna reception, and more complex propagation effects.Unlike images, each wireless measurement is a single complex number, and tracing antenna rays across directions strains memory and convergence.Source: §1 · p. 2
- NeWRF is a NeRF-based framework that integrates wireless propagation characteristics to learn wireless radiation scenes from sparse channel measurements.The paper presents it as the first such wireless channel prediction framework.Source: §1 · p. 2
- A direction-of-arrival-guided ray-casting scheme searches critical directions to improve convergence, while a ray-searching algorithm estimates direction of arrival for channel prediction at unvisited locations.These techniques address the large direction-search space of wireless antennas.Source: §1 · p. 2
- Inspecting NeWRF's learned model reveals wireless scenes' simple nature, which the authors say enables learning in complex room-scale spaces from sparse measurements.They describe the framework as a first step toward practical NeRF use in future wireless applications.Source: §1 · p. 2
2. Preliminaries
The preliminaries describe how propagation alters wireless signals and define the channel from transmitted and received signals. They also outline NeRF’s scene representation and volume-rendering process.
Wireless Channel
- The wireless channel h is the ratio of received to transmitted signal and characterizes distortions introduced by the environment.Signals can encounter attenuation and phase rotation, including changes from object interactions and multipath propagation.Source: §2 · p. 3
Neural Radiance Fields (NeRF)
- NeRF represents a scene with an MLP mapping 3D position and ray direction to RGB radiance and volume density.It queries sample points along camera rays to predict radiance and density.Source: §2 · p. 3
- NeRF translates volume density into transmittance and opacity, with opacity describing light absorbed or scattered by a volume block.The distance between consecutive samples is δ_i = ||p_i+1 − p_i||.Source: §2 · p. 3
- Volume rendering approximates a pixel value as a weighted sum of radiance at sample points along a ray.The generated image is compared with a ground-truth image to calculate the training loss and update model parameters.Source: §2 · p. 3
3. Related Work
Related work spans visible-light NeRF extensions and methods for wireless spatial prediction. The paper contrasts its approach with prior methods in their prediction settings, measurement density, and need for environment geometry.
Neural Radiance Fields
- Earlier NeRF extensions address large scenes, dynamic scenes, or dark images, while this work extends the formulation to radio-frequency radiation scenes.The cited extensions operate in the visible-light spectrum, which the paper distinguishes from wireless signals.Source: §3 · p. 3
Wireless channel prediction
- The paper distinguishes spatial prediction at unseen locations from temporal prediction at future timestamps and focuses on the spatial problem.Spatial prediction uses measurements at other locations; temporal prediction uses past observations.Source: §3 · p. 3
- Prior analytical prediction is limited to 2D layouts with carefully designed boundaries, while NeRF2 predicts direction-of-arrival heatmaps for fixed receivers.NeRF2 requires 178.1 measurements/ft3 on average, compared with 0.2 measurements/ft3 for the paper’s method.Source: §3 · p. 4
- A wireless ray-tracing surrogate requires CAD geometry, whereas this method learns the radiation field without environment geometry beyond the scene boundary.The paper says this removes the need for CAD models as replicas of the real world.Source: §3 · p. 4
4. NeWRF
NeWRF learns a wireless radiation scene from sparse channel measurements to predict channels at unvisited locations. Its design adapts NeRF with wireless propagation modeling, DoA-guided training, and ray-searching for inference.
4.1. Problem Description
- The prediction setting assumes one fixed transmitter and randomly placed receivers, with the goal of estimating channels throughout the environment from sparse measurements.When multiple transmitters coexist, the paper assumes receivers can connect to each in round-robin fashion and construct a separate scene for each.Source: §4.1 · p. 4
4.2. Neural Wireless Channel Synthesis
- For each receiver, NeWRF synthesizes a channel by querying an MLP at sampled points along backtraced rays and combining their weighted signal contributions.The MLP takes spatial and direction coordinates and predicts amplitude, phase, and volume density; reflected signals are represented as virtual transmitters.Source: §4.2 · p. 5, Figure 4 · p. 5
4.3. Improving the Convergence of NeWRF
- DoA-guided ray casting avoids an exhaustive angular search, reducing training time and memory use while enabling convergence.A 1° grid spans 64,800 directions, although only 10–20 ray directions contribute significantly at each receiver.Source: §4.3 · p. 5
4.4. Optimizing NeWRF
- Training uses coarse and fine MLPs, hierarchical sampling, and NMSE loss to match synthesized channels to measurements across varying channel scales.The model predicts real and imaginary signal parts because phase is not differentiable modulo 2π; 5–10 extra random-direction rays can remove floaters.Source: §4.4 · pp. 5–6
4.5. NeWRF’s Scene Representation
- Rather than placing high volume density on reflectors, the learned scene concentrates it at the transmitter and its virtual images, where backtraced rays intersect.This is one of multiple representations that fit the training set, but the model appears biased toward it and can still synthesize channels accurately from sparse measurements.Source: §4.5 · pp. 6–7, Figure 5 · p. 6
4.6. Predicting Channels at Unvisited Locations
- For unvisited locations, ray-searching uses training-set DoAs to identify and cluster potential virtual transmitters, then predicts their count and selects them by neighboring-receiver votes.The selected transmitters’ positions and the test receiver’s position determine the DoA; six nearest training receivers are used in practice.Source: §4.6 · p. 7
5. Evaluation
Evaluation uses simulated datasets across three 3D environments, with measurements split into training and test sets. It assesses channel prediction first with known DoA, then with DoA estimated by ray-searching.
5.1. Channel Prediction Results
- NeWRF outperforms all baseline methods in all three environments, while retaining more than 5 dB SNR in the larger, more complex office environment.Performance declines as environment size and complexity increase; the reported SNR remains sufficient for many wireless applications.Source: Figure 8 · p. 8, §5.1 · p. 8
5.2. Measurement Density of Training Set
- At 0.2 measurements/ft3, NeWRF predicts channels accurately—around 800× lower density than NeRF2 requires.The conference-room experiment evaluates training sets using 25%, 50%, 75%, and 100% of measurements on the same test set.Source: §5.2 · p. 9
5.3. Impact of Materials Types and Signal Frequencies
- NeWRF maintains high prediction accuracy across the tested reflector materials and signal frequencies, with a separate model trained for each condition.Material electrical properties and signal frequency affect wireless attenuation and phase rotation.Source: §5.3 · p. 9
5.4. Ray-searching Algorithm Performance
- Using ray-searching to estimate DoA reduces channel-estimation SNR by 3–5 dB relative to known DoA.The authors attribute the loss mainly to slight DoA discrepancies between test locations and their nearest neighbors; slightly higher measurement density may improve it.Source: §5.4 · p. 9
6. Limitations and Future work
The discussion identifies constraints in phase and DoA requirements and in scaling to more complex environments.
a) Relaxing phase requirement
- Inaccurate or unavailable phase limits NeWRF’s performance because the presented algorithm relies on narrowband channel measurements for training.The authors suggest wideband measurements as a potential way to relax this requirement.Source: §6 · p. 9
b) Relaxing DoA requirement
- Ray-casting requires DoA accuracy within one degree in the reported experiments, a resolution that may require a very large antenna array in practice.Further work is needed to relax this constraint.Source: §6 · p. 9
c) Extending to larger environments
- Although NeWRF outperforms baselines in the three studied environments, its performance decreases as environment complexity increases, and scaling needs further research.The authors suggest fusing visual and wireless radiation fields as one possible approach.Source: §6 · p. 9
7. Conclusion
NeWRF is presented as the first NeRF-based sparse-measurement framework for wireless channel prediction, addressing challenges in neural wireless-channel synthesis.
7. Conclusion
- NeWRF is presented as the first NeRF-based learning framework for wireless channel prediction with sparse measurements.The paper describes its work as a first step toward practical use of NeRF in future wireless applications.Source: §7 · p. 9
A. Implementation Details
Implementation uses the same MLP architecture for coarse and fine models, with positional encoding and ray sampling configured for each environment.
A. Implementation Details
- Training uses positional encoding with 10 frequencies for spatial coordinates and 4 for direction coordinates, and a batch size of 32.The positional encoding maps input coordinates to a higher-dimensional space to help model high-frequency variation.Source: A. Implementation Details · p. 12, Figure 10 · p. 12
- The NeWRF MLP has seven width-128 linear layers and a width-64 output layer, with a skip connection at the fourth layer and ReLU or tanh activations.The coarse and fine models share this architecture, and their losses are combined as a weighted sum.Source: A. Implementation Details · p. 12, Figure 11 · p. 12
- The model is trained with Adam at an initial learning rate of 5 × 10^-4 and typically converges in about 100k iterations on one NVIDIA A100 GPU.A ReduceLROnPlateau scheduler uses patience 10 and factor 0.9.Source: A. Implementation Details · p. 12
- Ray sampling uses environment-specific ranges and resolutions, doubling both sampling stages for the more complex office environment.The conference room and bedroom use 128 coarse and 128 hierarchical samples per ray; the office doubles both stages.Source: A. Implementation Details · p. 12
B. Subcarrier Channel Prediction
NeWRF predicts nearby subcarrier channels accurately from a model trained on a single subcarrier, without further fine-tuning. Its channel synthesis separates learned radiated signals from frequency-dependent propagation effects.
B. Subcarrier Channel Prediction
- In OFDM, receivers estimate channels for each subcarrier; commercial WiFi chips typically report CSI for 52 of 64 subcarriers in a 20 MHz channel.The 64 subcarriers are spaced 312.5 kHz apart.Source: B. Subcarrier Channel Prediction · p. 12
- A model trained on one subcarrier predicts nearby subcarrier channels accurately without fine-tuning, including highly frequency-selective channels.Predictions align well with ground truth across five receiver locations.Source: B. Subcarrier Channel Prediction · pp. 12–13, Figure 13 · p. 13
- NeWRF decouples reflection-related attenuation and phase rotation from free-space propagation, modeling frequency-dependent phase shifts and multipath interference in channel synthesis.The model predicts the radiated signal, which is nearly unchanged across subcarrier frequencies.Source: B. Subcarrier Channel Prediction · p. 13
C. Impact of Noise
NeWRF’s prediction accuracy depends on measurement noise, but training with measurements having different additive biases maintains good performance even at very low SNRs. The authors attribute this robustness to implicit averaging during training.
C. Impact of Noise
- Training with measurements containing different additive noises maintains good performance even at very low SNRs.The authors attribute the effect to the model implicitly averaging measurements to reduce bias.Source: C. Impact of Noise · p. 13
- With a fixed noise bias added to training channels, prediction accuracy decreases as channel SNR decreases.Channel SNR is the ratio between channel power and added noise power.Source: C. Impact of Noise · p. 13
- More than 20 dB channel SNR is typical in wireless systems for reliable communication, and NeWRF performs well at such SNRs.Source: C. Impact of Noise · p. 13