Research map / Image generation / Image restoration

Physics Unrolled Neural Operator for Wireless Field Modeling

NeurIPS 2026 · Image generation › Image restoration · approach: Diffusion priors for denoising and deblurring · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes PU-HNO, a three-stage physics-unrolled neural operator that refines low-fidelity ray-traced radio maps into high-fidelity predictions, trained only on noisy intermediate-fidelity labels, with a zero-shot denoising theorem.

Problem

Predicting high-fidelity indoor radio maps (RSS) from cheap low-fidelity ray tracing when available training labels are intermediate-fidelity Monte Carlo ray-tracing outputs with residual noise and no high-fidelity labels.

Approach

A cascade of three residual neural operators, each targeting one propagation mechanism (specular reflection, diffraction, scattering), with stage-wise supervision, geometry-conditioned operators, an adaptive residual gate, and wireless-aware losses (confidence-weighted Huber, Sobolev gradient, transport-feature alignment); theoretically, conditionally unbiased label noise lets the model outperform

Key results

• PU-HNO achieves RMSE 3.471 dB, SSIM 0.928, and Outage F1 0.820 against held-out HF-GT, versus 14.012 dB RMSE for the training labels and 3.851 dB for the best U-Net baseline.
• Ablation: removing semantic/material priors drops Outage F1 from 0.820 to 0.001; the staged architecture raises Outage F1 from 0.206 (specular only) to 0.399 (+diffraction) to 0.820 (+scattering).
• Access-point placement on a 10,000 m^2 floorplan yields 92 Mbps throughput for PU-HNO versus 16 Mbps for CNN despite only a 2.7% SSIM difference.

Surprising findings

• Several image-quality-competitive baselines (e.g., U-Net SSIM 0.919) have very low Outage F1 (0.18), showing image metrics miss coverage-relevant structure.
• Heteroscedastic label noise slightly improves RMSE (3.84 to 3.62 dB), suggesting a regularization effect.

Limitations

• Depends on accurate geometry, obstacle, and material priors; systematic bias in labels (geometry/distance-dependent) causes large failure (RMSE 12.06 dB).
• Only RSS at a single 5.5 GHz frequency, indoor, single transmitter, scalar output; no phase or CSI.
• All supervision and evaluation are from ray tracing; no sim-to-real validation with measured data.

Experimental setup

Datasets: Procedural indoor RSS dataset generated with Sionna RT (LF 10^4 rays, IF-GT 10^6 rays, HF-GT 10^8 rays)

Benchmarks: PU-HNO Procedural Indoor RSS Dataset

Metrics: Outage F1 Fading Ratio Tail Error (5%) MCESE RMSE MAE SSIM LPIPS throughput (access-point placement)

Baselines: CNN U-Net ResNet ViT NeRF^2 RadioUNet GeneRT WiGATr FNO TFNO UNO SFNO

Builds on: Fourier neural operator wavelet neural operator U-Net RadioUNet Sionna RT

Closest papers by meaning

• Radio Frequency Ray Tracing with Neural Object Representation for Enhanced RF Modeling (CVPR 2025 · similarity 0.86)
• NeWRF: A Deep Learning Framework for Wireless Radiation Field Reconstruction and Channel Prediction (ICML 2024 · similarity 0.85)
• Differentiable and Learnable Wireless Simulation with Geometric Transformers (ICLR 2025 · similarity 0.85)
• RxGS: Receiver-Generalizable 3D Gaussian Splatting for Radio-Frequency Data Synthesis (NeurIPS 2026 · similarity 0.83)
• NEAR: Neural Electromagnetic Array Response (ICML 2025 · similarity 0.82)
• Model-Driven Deep Neural Network for Enhanced AoA Estimation Using 5G gNB (AAAI 2024 · similarity 0.82)