Research map / Probabilistic and scientific modeling / Molecules and proteins

Physics-Informed Generative Modeling of Wireless Channels

ICML 2025 · Probabilistic and scientific modeling › Molecules and proteins · approach: Machine learning for physical simulation · 1 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.

Physics-informed generative models for wireless channels that learn site-specific distributions of sparse physical channel parameters from compressed, noisy AP observations, using prescribed SBGM (CSVAE/CSGMM) and an implicit AmbientGAN variant.

Problem

Learning site-specific wireless channel distributions from compressed, noisy observations without high-quality ground-truth training data, while ensuring generalizability across system configurations and physical interpretability.

Approach

Combine the compressibility of wireless channels in a physically interpretable dictionary (delay/Doppler/angle grids) with sparse Bayesian generative modeling (CSVAE and CSGMM), so the model learns a distribution over sparse path-gain vectors from compressed observations; the zero-mean diagonal-covariance structure matches the Toeplitz structure of physical channel moments.

Key results

• On the modified 3GPP SIMO setup, CSVAE best matches the ground-truth power angular profile and CSGMM closely matches the angular spread histogram, while AmbientGAN significantly overestimates angular spread.
• On QuaDRiGa 5G-Urban, CSVAE and CSGMM achieve the best autoencoder-reconstruction nMSE and cosine similarity; CSGMM approaches ground-truth performance even for small N_t.
• On 5G-Rural with increasing Doppler/delay bounds, AmbientGAN for h degrades significantly while CSGMM and CSVAE maintain strong performance; M-SBL performs worst.

Measured effects reported in the paper

Surprising findings

• AmbientGAN trained to output channels h performs significantly worse than all other methods on 5G-Urban.
• M-SBL, a single-component Gaussian fit, performs consistently well on 5G-Urban, suggesting the true sparse-parameter distribution there is close to Gaussian.

Limitations

• Cannot model non-stationary channel generation (WSSUS / far-field assumption).
• Off-grid mismatch due to discretization of the parameter space.
• OFDM pilot patterns: regular pilot patterns degrade compressed-sensing-based training.

Experimental setup

Datasets: modified 3GPP spatial channel model QuaDRiGa 5G-Urban QuaDRiGa 5G-Rural DeepMIMO

Benchmarks: modified 3GPP SIMO QuaDRiGa 5G-Urban QuaDRiGa 5G-Rural DeepMIMO

Metrics: power angular profile channel-wise angular spread nMSE cosine similarity

Baselines: AmbientGAN (for s) AmbientGAN (for h) M-SBL ground-truth QuaDRiGa channels

Builds on: sparse Bayesian learning AmbientGAN variational autoencoder Wasserstein GAN with gradient penalty compressive sensing generative priors (SBGM)

Closest papers by meaning

• Sparse Bayesian Generative Modeling for Compressive Sensing (NeurIPS 2024 · similarity 0.85)
• Stochastic Neural Ray Tracing for Radio Frequency Channel Modeling (ICML 2026 · similarity 0.84)
• Modeling Sparse Deviations for Compressed Sensing using Generative Models (ICML 2018 · similarity 0.84)
• NeWRF: A Deep Learning Framework for Wireless Radiation Field Reconstruction and Channel Prediction (ICML 2024 · similarity 0.82)
• Sparse Gaussian Processes: Structured Approximations and Power-EP Revisited (NeurIPS 2025 · similarity 0.82)
• Physics Unrolled Neural Operator for Wireless Field Modeling (NeurIPS 2026 · similarity 0.81)