pith. sign in

arxiv: 2605.15471 · v1 · pith:TXXTO7CInew · submitted 2026-05-14 · 📡 eess.SP

CITYMPC: A Large-Scale Physics-Informed Benchmark and Tool for Generative Complete Multipath Wireless Channel Modeling

Ashwin Natraj Arun , David R. Nickel , Yaguang Zhang , Yunchou Xing , Jie Chen , Amitava Ghosh , Christopher Brinton , David J. Love
show 1 more author
James V. Krogmeier
This is my paper

Pith reviewed 2026-05-19 14:30 UTC · model grok-4.3

classification 📡 eess.SP
keywords multipath channel modelingconditional variational autoencoderray tracingurban wireless channelsgenerative modelsscene-conditioned predictiondistribution shiftwireless benchmark dataset
0
0 comments X

The pith

CITYMPC generates complete multipath wireless channel parameters from point-of-view imagery and terrain height maps alone.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces CITYMPC as a conditional variational autoencoder that outputs the full set of multipath component parameters including complex gain, propagation delay, and angles of arrival and departure. It relies solely on point-of-view imagery and terrain height maps rather than full three-dimensional scene models during use. The system was trained and tested on ray-traced data from five urban areas covering over 427,000 links and reaches mean absolute errors of 1.29 dB in received power and 7.25 ns in first-path delay relative to ray tracing. The authors also release the dataset and examine how performance changes when the model is applied to cities outside the training set. This setup offers a way to create environment-specific channel models at lower computational cost than traditional ray tracing.

Core claim

CITYMPC is a conditional variational autoencoder that predicts the complete per-path multipath component parameter set, including complex channel gain, propagation delay, azimuth and elevation angles of departure and arrival, directly from point-of-view imagery and terrain height maps, achieving received power mean absolute error of 1.29 dB and first-path delay mean absolute error of 7.25 ns against ray-tracing ground truth across 427,397 links in five urban environments without requiring explicit three-dimensional scene geometry at inference time.

What carries the argument

Conditional variational autoencoder conditioned on point-of-view imagery and terrain height maps to output the full multipath component parameter set.

If this is right

  • Supports large-scale environment-aware channel generation for wireless network planning without running ray tracing on every new scene.
  • Enables release of a multi-city ray-traced dataset as a public benchmark for future scene-conditioned models.
  • Allows quantitative measurement of cross-city distribution shift to guide model robustness improvements.
  • Provides a generative framework that can sample varied channel realizations consistent with a given visual and terrain input.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method could be paired with publicly available mapping imagery to produce channel models for cities lacking detailed 3D reconstructions.
  • Extension to sequences of images might support modeling of time-varying channels caused by moving vehicles or pedestrians.
  • Similar conditioning on visual and elevation data could be tested in non-urban settings such as suburban or indoor environments to broaden applicability.

Load-bearing premise

That point-of-view imagery and terrain height maps alone contain enough information to accurately reconstruct the full set of multipath component parameters that would otherwise require explicit 3D scene geometry.

What would settle it

Compare CITYMPC predictions against real-world channel measurements collected in a sixth urban area never seen during training and check whether the received power and delay errors remain near 1.29 dB and 7.25 ns.

Figures

Figures reproduced from arXiv: 2605.15471 by Amitava Ghosh, Ashwin Natraj Arun, Christopher Brinton, David J. Love, David R. Nickel, James V. Krogmeier, Jie Chen, Yaguang Zhang, Yunchou Xing.

Figure 1
Figure 1. Figure 1: Top: pictorial representation of multipath propagation. Bottom: sam￾ple Sionna RT datapoint. Related work. Stochastic channel models such as 3GPP TR 38.901 [8] and NYUSIM [11] provide closed-form parametric dis￾tributions but lack site-specific conditioning. Ray tracing tools such as Sionna RT [9] and Wireless InSite [12] produce physically accurate channels at high computational cost, requiring detailed 3… view at source ↗
Figure 2
Figure 2. Figure 2: Conditioning inputs for the Tx/Rx link shown in Fig. 1. The global height map (left) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: System diagram for CITYMPC multipath channel generation. During training, samples are drawn from the latent space generated by the posterior encoder. For inference, the prior network, conditioned only on scene information c, is used for generating a sample from the latent space. Inputs and outputs. Let c = {Itx ∈ R 12×128×128 , Irx ∈ R 12×128×128 , Ig ∈ R 1×128×128 , s ∈ R 6} denote the conditioning inform… view at source ↗
Figure 4
Figure 4. Figure 4: A channel realization generated by CITYMPC for a single Austin Tx-Rx link. Columns show from left to right: CIR amplitude |h(τ )| in dB versus absolute delay τ (ns), AoD azimuth polar plot, AoD elevation polar plot, AoA azimuth polar plot, and AoA elevation polar plot. Ground truth is shown in blue and predicted paths in orange in all panels. Quantitative results. For each city we run a seed sweep and repo… view at source ↗
Figure 5
Figure 5. Figure 5: Spatial maps of total received power (dB) across all Rx locations for a fixed Tx position [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Empirical CDFs of ground-truth (blue) and generated (orange) channel parameters for [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Cross-city transfer matrices for received power MAE, ToF MAE, and presence F1. Rows [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Two additional independent channel realizations generated by [PITH_FULL_IMAGE:figures/full_fig_p024_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Empirical CDFs of ground-truth (blue) and generated (orange) channel parameters for [PITH_FULL_IMAGE:figures/full_fig_p025_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Empirical CDFs of ground-truth (blue) and generated (orange) channel parameters for [PITH_FULL_IMAGE:figures/full_fig_p025_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Empirical CDFs of ground-truth (blue) and generated (orange) channel parameters for [PITH_FULL_IMAGE:figures/full_fig_p026_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Empirical CDFs of ground-truth (blue) and generated (orange) channel parameters for [PITH_FULL_IMAGE:figures/full_fig_p026_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Cross-city transfer matrices for average delay MAE, average AoD azimuth MAE, average [PITH_FULL_IMAGE:figures/full_fig_p027_13.png] view at source ↗
read the original abstract

Multipath wireless channels are fully characterized by multipath components (MPCs), including complex channel gain, propagation delay, angle of departure (AoD) and angle of arrival (AoA) in azimuth and elevation. Generating these parameters with the fidelity of ray tracing (RT) remains an open problem. Existing methods either incur the computational cost of RT or require explicit 3D scene geometry at inference. We present CITYMPC, a conditional variational autoencoder (cVAE) that predicts the complete per-path MPC parameter set from point-of-view imagery and terrain height maps alone, achieving environment-aware channel generation without access to any three-dimensional scene geometry at inference. Trained and evaluated across five urban environments spanning 427,397 links, CITYMPC matches RT ground truth to within 1.29 dB received power mean absolute error (MAE) and 7.25 ns $\tau_0$ MAE. CITYMPC is a generative channel modeling framework and reproducible benchmark, released together with a large-scale multi-city ray-traced dataset to accelerate future scene-conditioned channel modeling research. We further analyze cross-city distribution shift to characterize the per-city diversity of the benchmark.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces CITYMPC, a conditional variational autoencoder (cVAE) that generates the full set of multipath component (MPC) parameters—including complex gain, delay, AoD, and AoA in azimuth and elevation—from point-of-view imagery and terrain height maps alone. It claims to match ray-tracing ground truth to within 1.29 dB received-power MAE and 7.25 ns delay MAE across 427,397 links in five urban environments, while releasing a large-scale multi-city ray-traced dataset and benchmark for scene-conditioned channel modeling. The work also includes cross-city distribution-shift analysis.

Significance. If the performance claims prove robust, the contribution would be significant: it offers a practical route to environment-aware generative channel modeling that avoids both the runtime cost of ray tracing and the need for explicit 3-D geometry at inference time. The scale of the released dataset and the explicit treatment of cross-city shift are concrete assets that could accelerate follow-on research in wireless propagation modeling.

major comments (2)
  1. [§3] §3 (Model Architecture and Training): The manuscript provides no description of the cVAE encoder/decoder architecture, the precise form of the evidence lower bound, the weighting of reconstruction versus KL terms, the handling of variable-length path sets, or the train/validation/test splits. Without these details the reported aggregate MAEs cannot be verified as free of overfitting or leakage, directly undermining the central claim that the model generalizes from imagery and terrain maps alone.
  2. [§4.3] §4.3 (Per-path and Cross-city Results): The 1.29 dB power and 7.25 ns delay MAEs are reported only in aggregate; no per-path error statistics are given for angle parameters, nor is there a breakdown of performance on paths whose reflectors lie outside the camera FOV or below terrain-map resolution. Urban ray-tracing ground truth routinely contains such occluded paths; if the cVAE cannot recover them, the per-path fidelity claim fails even when aggregate statistics appear acceptable.
minor comments (2)
  1. [Abstract] The title and abstract describe the method as “physics-informed,” yet the presented cVAE is a standard conditional generative model with no explicit physics constraints or ray-tracing priors inside the network; a clarifying sentence would avoid misinterpretation.
  2. [§5] Figure captions and axis labels in the cross-city shift plots should explicitly state the number of links per city and the exact metric used for the shift quantification.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments, which have helped us strengthen the reproducibility and completeness of the manuscript. We address each major comment below and have incorporated revisions to provide the requested details and analyses.

read point-by-point responses

  1. Referee: [§3] §3 (Model Architecture and Training): The manuscript provides no description of the cVAE encoder/decoder architecture, the precise form of the evidence lower bound, the weighting of reconstruction versus KL terms, the handling of variable-length path sets, or the train/validation/test splits. Without these details the reported aggregate MAEs cannot be verified as free of overfitting or leakage, directly undermining the central claim that the model generalizes from imagery and terrain maps alone.

    Authors: We agree that these implementation details are necessary for full reproducibility and to substantiate the generalization claims. In the revised manuscript we have expanded Section 3 with a complete description of the cVAE, including the encoder and decoder network architectures (layer counts, dimensions, and activations), the exact ELBO objective, the specific weighting coefficients between reconstruction and KL terms together with their selection rationale, the padding-and-masking strategy used to accommodate variable-length MPC sets, and the precise train/validation/test split ratios with city-wise stratification to avoid leakage. These additions allow independent verification that the reported MAEs reflect genuine generalization rather than overfitting. revision: yes

  2. Referee: [§4.3] §4.3 (Per-path and Cross-city Results): The 1.29 dB power and 7.25 ns delay MAEs are reported only in aggregate; no per-path error statistics are given for angle parameters, nor is there a breakdown of performance on paths whose reflectors lie outside the camera FOV or below terrain-map resolution. Urban ray-tracing ground truth routinely contains such occluded paths; if the cVAE cannot recover them, the per-path fidelity claim fails even when aggregate statistics appear acceptable.

    Authors: We acknowledge that aggregate metrics alone leave open questions about per-path fidelity, especially for angles and occluded paths. The revised manuscript now reports per-path MAE statistics for AoD and AoA (azimuth and elevation). We also add a stratified breakdown separating paths whose dominant reflectors lie inside versus outside the camera FOV or terrain-map coverage. The results show expected degradation on fully occluded paths, yet the generative model still produces aggregate channel statistics (power and delay) that remain close to ray-tracing ground truth. We have clarified in the text that the central claims concern environment-aware channel statistics rather than exact per-path recovery for every individual MPC. revision: yes

Circularity Check

0 steps flagged

No significant circularity: standard supervised generative model trained on external RT labels

full rationale

The paper trains a cVAE on ray-traced ground-truth MPC parameters (derived from explicit 3D geometry during data generation) to learn a conditional mapping from POV imagery plus terrain height maps. At inference the model applies only the learned parameters to new imagery/terrain inputs and outputs the per-path set; this is a conventional supervised learning pipeline whose outputs are not algebraically or definitionally identical to the training inputs. No self-definitional equations, fitted-input predictions, load-bearing self-citations, imported uniqueness theorems, or smuggled ansatzes appear in the abstract or described method. The central claim therefore remains independent of the target result and receives the default non-circularity finding.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Based on abstract only the ledger is necessarily incomplete; the approach rests on standard variational inference assumptions plus the domain premise that visual and height inputs suffice for MPC prediction.

axioms (2)
  • standard math Variational inference provides a tractable approximation to the posterior distribution in the conditional VAE
    Core to any VAE training procedure
  • domain assumption Imagery and terrain height maps encode sufficient information about propagation paths to allow accurate MPC parameter prediction
    Load-bearing premise for removing explicit 3D geometry requirement

pith-pipeline@v0.9.0 · 5774 in / 1410 out tokens · 48643 ms · 2026-05-19T14:30:31.088545+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

56 extracted references · 56 canonical work pages · 5 internal anchors

  1. [1]

    Recommendation M.2160: Framework and Overall Objectives of the Future Devel- opment of IMT for 2030 and Beyond, 2023

    ITU-R. Recommendation M.2160: Framework and Overall Objectives of the Future Devel- opment of IMT for 2030 and Beyond, 2023. URL https://www.itu.int/rec/R-REC-M. 2160/en. Accessed: 2026-01-24

    work page 2030

  2. [2]

    6G Roadmap for Vertical Industries, 2023

    Next G Alliance. 6G Roadmap for Vertical Industries, 2023. URL https://nextgalliance. org/white_papers/6g-roadmap-vertical-industries/ . ATIS, Accessed: 2026-01-24

    work page 2023

  3. [3]

    Brinton, Mung Chiang, Kwang Taik Kim, David J

    Christopher G. Brinton, Mung Chiang, Kwang Taik Kim, David J. Love, Michael Beesley, Morris Repeta, John Roese, Per Beming, Erik Ekudden, Clara Li, Geng Wu, Nishant Batra, Amitava Ghosh, V olker Ziegler, Tingfang Ji, Rajat Prakash, and John Smee. Key Focus Areas and Enabling Technologies for 6G. IEEE Communications Magazine, 63(3):84–91, 2025. doi: 10.110...

    work page doi:10.1109/mcom.002.2400150 2025

  4. [4]

    Hwanjin Kim, Junil Choi, and David J. Love. Machine-Learning Techniques for Wireless Channel Prediction: Insights and Practical Guidance. IEEE Wireless Communications, pages 1–8, 2026. doi: 10.1109/MWC.2026.3678210

    work page doi:10.1109/mwc.2026.3678210 2026

  5. [5]

    Radio Propagation Measurements and Channel Modeling

    Theodore S Rappaport, Kate A Remley, Camillo Gentile, Andreas F Molisch, and Alenka Zaji´c. Radio Propagation Measurements and Channel Modeling. Cambridge University Press, 2022

    work page 2022

  6. [6]

    Samimi and Theodore S

    Mathew K. Samimi and Theodore S. Rappaport. 3-D Millimeter-Wave Statistical Channel Model for 5G Wireless System Design. IEEE Transactions on Microwave Theory and Techniques, 64 (7):2207–2225, 2016. doi: 10.1109/TMTT.2016.2574851

    work page doi:10.1109/tmtt.2016.2574851 2016

  7. [7]

    Fundamentals of Wireless Communication

    David Tse and Pramod Viswanath. Fundamentals of Wireless Communication. Cambridge University Press, 2005

    work page 2005

  8. [8]

    Study on Channel Model for Frequencies from 0.5 to 100 GHz, 2026

    3rd Generation Partnership Project (3GPP). Study on Channel Model for Frequencies from 0.5 to 100 GHz, 2026. URL https://www.3gpp.org/ftp/Specs/archive/38_series/ 38.901/38901-j20.zip. TR 38.901 v19.1.0, accessed: 2026-01-24

    work page 2026

  9. [9]

    Sionna, 2022

    Jakob Hoydis, Sebastian Cammerer, Fayçal Ait Aoudia, Merlin Nimier-David, Lorenzo Maggi, Guillermo Marcus, Avinash Vem, and Alexander Keller. Sionna, 2022. https://nvlabs. github.io/sionna/

    work page 2022

  10. [10]

    Sionna RT: Differentiable Ray Tracing for Radio Propagation Modeling

    Jakob Hoydis, Faycal Ait Aoudia, Sebastian Cammerer, Merlin Nimier-David, Nikolaus Binder, Guillermo Marcus, and Alexander Keller. Sionna RT: Differentiable Ray Tracing for Radio Propagation Modeling. In 2023 IEEE Globecom Workshops (GC Wkshps), pages 317–321,

    work page 2023

  11. [11]

    doi: 10.1109/GCWkshps58843.2023.10465179

    work page doi:10.1109/gcwkshps58843.2023.10465179 2023

  12. [12]

    MacCartney, and Theodore S

    Shu Sun, George R. MacCartney, and Theodore S. Rappaport. A novel millimeter-wave channel simulator and applications for 5g wireless communications. In 2017 IEEE International Conference on Communications (ICC), pages 1–7, 2017. doi: 10.1109/ICC.2017.7996792

    work page doi:10.1109/icc.2017.7996792 2017

  13. [13]

    Wireless InSite: 3D Wireless Prediction Software, 2021

    Remcom. Wireless InSite: 3D Wireless Prediction Software, 2021

    work page 2021

  14. [14]

    ChannelGAN: Deep Learning-Based Channel Modeling and Generating

    Han Xiao, Wenqiang Tian, Wendong Liu, and Jia Shen. ChannelGAN: Deep Learning-Based Channel Modeling and Generating. IEEE Wireless Communications Letters, 11(3):650–654,

    work page

  15. [15]

    doi: 10.1109/LWC.2021.3140102

    work page doi:10.1109/lwc.2021.3140102 2021

  16. [16]

    Genera- tive Diffusion Models for Radio Wireless Channel Modelling and Sampling

    Ushnish Sengupta, Chinkuo Jao, Alberto Bernacchia, Sattar Vakili, and Da-shan Shiu. Genera- tive Diffusion Models for Radio Wireless Channel Modelling and Sampling. In GLOBECOM 2023 - 2023 IEEE Global Communications Conference, pages 4779–4784, 2023. doi: 10.1109/GLOBECOM54140.2023.10437154

    work page doi:10.1109/globecom54140.2023.10437154 2023

  17. [17]

    Schaefer

    Muah Kim, Rick Fritschek, and Rafael F. Schaefer. Diffusion models for accurate channel distribution generation, 2024. URLhttps://arxiv.org/abs/2309.10505

    work page arXiv 2024

  18. [18]

    Taekyun Lee, Juseong Park, Hyeji Kim, and Jeffrey G. Andrews. Generating high dimen- sional user-specific wireless channels using diffusion models. IEEE Transactions on Wireless Communications, 25:2907–2921, 2026. doi: 10.1109/TWC.2025.3600286. 11

    work page doi:10.1109/twc.2025.3600286 2026

  19. [19]

    WiNeRT: Towards Neural Ray Tracing for Wireless Channel Modelling and Differ- entiable Simulations

    Tribhuvanesh Orekondy, Pratik Kumar, Shreya Kadambi, Hao Ye, Joseph Soriaga, and Arash Behboodi. WiNeRT: Towards Neural Ray Tracing for Wireless Channel Modelling and Differ- entiable Simulations. In The Eleventh International Conference on Learning Representations,

    work page

  20. [20]

    URLhttps://openreview.net/forum?id=tPKKXeW33YU

    work page

  21. [21]

    GeNeRT: A Physics-Informed Approach to Intelligent Wireless Channel Modeling via Generalizable Neural Ray Tracing, 2025

    Kejia Bian, Meixia Tao, Shu Sun, and Jun Yu. GeNeRT: A Physics-Informed Approach to Intelligent Wireless Channel Modeling via Generalizable Neural Ray Tracing, 2025. URL https://arxiv.org/abs/2506.18295

    work page arXiv 2025

  22. [22]

    Xiucheng Wang, Qiming Zhang, Nan Cheng, Junting Chen, Zezhong Zhang, Zan Li, Shuguang Cui, and Xuemin Shen. RadioDiff-3D: A 3D× 3D Radio Map Dataset and Generative Diffusion Based Benchmark for 6G Environment-Aware Communication.IEEE Transactions on Network Science and Engineering, 13:3773–3789, 2026. doi: 10.1109/TNSE.2025.3590545

    work page doi:10.1109/tnse.2025.3590545 2026

  23. [23]

    Agrawal, Richard Newcombe, and Ahmed Alkhateeb

    Shuaifeng Jiang, Qi Qu, Xiaqing Pan, Abhishek K. Agrawal, Richard Newcombe, and Ahmed Alkhateeb. Learnable wireless digital twins: Reconstructing electromagnetic field with neural representations. IEEE Open Journal of the Communications Society, 6:1568–1590, 2025. doi: 10.1109/OJCOMS.2025.3535959

    work page doi:10.1109/ojcoms.2025.3535959 2025

  24. [24]

    Juseong Park, Taekyun Lee, Yunchou Xing, Jie Chen, Amitava Ghosh, and Jeffrey G. Andrews. Learning ray-tracing for radio propagation via cross-attention-based diffusion models. In 2025 59th Asilomar Conference on Signals, Systems, and Computers, pages 1210–1214, 2025. doi: 10.1109/IEEECONF67917.2025.11443807

    work page doi:10.1109/ieeeconf67917.2025.11443807 2025

  25. [25]

    RadioUNet: Fast Ra- dio Map Estimation With Convolutional Neural Networks

    Ron Levie, Ça ˘gkan Yapar, Gitta Kutyniok, and Giuseppe Caire. RadioUNet: Fast Ra- dio Map Estimation With Convolutional Neural Networks. IEEE Transactions on Wireless Communications, 20(6):4001–4015, 2021. doi: 10.1109/TWC.2021.3054977

    work page doi:10.1109/twc.2021.3054977 2021

  26. [26]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors,Computer Vision – ECCV 2020, pages 213–229, Cham, 2020. Springer International Publishing. ISBN 978-3-030-58452-8

    work page 2020

  27. [27]

    Multi-Task Learning Using Uncertainty to Weigh Losses for Scene Geometry and Semantics

    Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, 2018. URL https://arxiv.org/abs/ 1705.07115

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  28. [28]

    Alkhateeb

    A. Alkhateeb. DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications. In Proc. of Information Theory and Applications Workshop (ITA), pages 1–8, San Diego, CA, Feb 2019

    work page 2019

  29. [29]

    Rappaport, Shu Sun, Rimma Mayzus, Hang Zhao, Yaniv Azar, Kevin Wang, George N

    Theodore S. Rappaport, Shu Sun, Rimma Mayzus, Hang Zhao, Yaniv Azar, Kevin Wang, George N. Wong, Jocelyn K. Schulz, Mathew Samimi, and Felix Gutierrez. Millimeter wave mobile communications for 5g cellular: It will work! IEEE Access, 1:335–349, 2013. doi: 10.1109/ACCESS.2013.2260813

    work page doi:10.1109/access.2013.2260813 2013

  30. [30]

    Mitsuba 3 renderer,

    Wenzel Jakob, Sébastien Speierer, Nicolas Roussel, Merlin Nimier-David, Delio Vicini, Tizian Zeltner, Baptiste Nicolet, Miguel Crespo, Vincent Leroy, and Ziyi Zhang. Mitsuba 3 renderer,

    work page

  31. [31]

    https://mitsuba-renderer.org

    work page

  32. [32]

    Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T

    Matthew Tancik, Pratul P. Srinivasan, Ben Mildenhall, Sara Fridovich-Keil, Nithin Raghavan, Utkarsh Singhal, Ravi Ramamoorthi, Jonathan T. Barron, and Ren Ng. Fourier features let networks learn high frequency functions in low dimensional domains. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS ’20, Red H...

    work page 2020

  33. [33]

    Auto-Encoding Variational Bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes, 2022. URL https: //arxiv.org/abs/1312.6114

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  34. [34]

    Learning structured output representation using deep conditional generative models

    Kihyuk Sohn, Honglak Lee, and Xinchen Yan. Learning structured output representation using deep conditional generative models. In C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett, editors, Advances in Neural Information Processing Systems, volume 28. Curran Associates, Inc., 2015. URL https://proceedings.neurips.cc/paper_files/paper/ 2015/file/...

    work page 2015

  35. [35]

    Channel vision transformers: An image is worth 1 x 16 x 16 words

    Yujia Bao, Srinivasan Sivanandan, and Theofanis Karaletsos. Channel vision transformers: An image is worth 1 x 16 x 16 words. In The Twelfth International Conference on Learning Representations, 2024. URLhttps://openreview.net/forum?id=CK5Hfb5hBG

    work page 2024

  36. [36]

    VAE with a VampPrior

    Jakub M. Tomczak and Max Welling. Vae with a vampprior, 2018. URL https://arxiv. org/abs/1705.07120

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  37. [37]

    Kingma, Tim Salimans, Rafal Jozefowicz, Xi Chen, Ilya Sutskever, and Max Welling

    Diederik P. Kingma, Tim Salimans, Rafal Jozefowicz, Xi Chen, Ilya Sutskever, and Max Welling. Improved variational inference with inverse autoregressive flow. In Proceedings of the 30th International Conference on Neural Information Processing Systems, NIPS’16, page 4743–4751, Red Hook, NY , USA, 2016. Curran Associates Inc. ISBN 9781510838819

    work page 2016

  38. [38]

    Generating sentences from a continuous space

    Samuel R Bowman, Luke Vilnis, Oriol Vinyals, Andrew M Dai, Rafal Jozefowicz, and Samy Bengio. Generating sentences from a continuous space. In 20th SIGNLL Conference on Computational Natural Language Learning, CoNLL 2016, pages 10–21. Association for Com- putational Linguistics (ACL), 2016

    work page 2016

  39. [39]

    Decoupled Weight Decay Regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. URL https: //arxiv.org/abs/1711.05101

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  40. [40]

    Lukas Kirchdorfer, Cathrin Elich, Simon Kutsche, Heiner Stuckenschmidt, Lukas Schott, and Jan M. Köhler. Analytical uncertainty-based loss weighting in multi-task learning. International Journal of Computer Vision, 2025. doi: 10.1007/s11263-025-02625-x. Preprint: arXiv:2408.07985

    work page doi:10.1007/s11263-025-02625-x 2025

  41. [41]

    Auxiliary Tasks in Multi-task Learning

    Lukas Liebel and Marco Körner. Auxiliary tasks in multi-task learning, 2018. URL https: //arxiv.org/abs/1805.06334

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  42. [42]

    Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio

    Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Proceedings of the 28th International Conference on Neural Information Processing Systems - V olume2, NIPS’14, page 2672–2680, Cambridge, MA, USA, 2014. MIT Press

    work page 2014

  43. [43]

    Denoising diffusion probabilistic models, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models, 2020

    work page 2020

  44. [44]

    Tribhuvanesh Orekondy, Arash Behboodi, and Joseph B. Soriaga. MIMO-GAN: Generative MIMO Channel Modeling, 2022

    work page 2022

  45. [45]

    Hardware-conditioned generative channel modeling: A diffusion-based approach for location and hardware-aware wireless dataset synthesis

    Nitish Deshpande, Sanjay Ganapathy, and Viraj Shah. Hardware-conditioned generative channel modeling: A diffusion-based approach for location and hardware-aware wireless dataset synthesis. In Open Conference of AI Agents for Science 2025, 2025. URL https: //openreview.net/forum?id=ExGHHgTM2p

    work page 2025

  46. [46]

    Physics-informed generative modeling of wireless channels, 2025

    Benedikt Böck, Andreas Oeldemann, Timo Mayer, Francesco Rossetto, and Wolfgang Utschick. Physics-informed generative modeling of wireless channels, 2025. URL https: //openreview.net/forum?id=FFJFT93oa7

    work page 2025

  47. [47]

    Love, and Christopher G

    Satyavrat Wagle, Akshay Malhotra, Shahab Hamidi-Rad, Aditya Sant, David J. Love, and Christopher G. Brinton. Physics-informed generative approaches for wireless channel modeling,

    work page

  48. [48]

    URLhttps://arxiv.org/abs/2503.05988

    work page arXiv

  49. [49]

    A geometry-based stochastic wireless channel model using channel images,

    Seongjoon Kang. A geometry-based stochastic wireless channel model using channel images,

    work page

  50. [50]

    URLhttps://arxiv.org/abs/2312.06637

    work page arXiv

  51. [51]

    NeRF2: Neural Radio-Frequency Radiance Fields

    Xiaopeng Zhao, Zhenlin An, Qingrui Pan, and Lei Yang. NeRF2: Neural Radio-Frequency Radiance Fields. In Proceedings of the 29th Annual International Conference on Mobile Computing and Networking, ACM MobiCom ’23, New York, NY , USA, 2023. Association for Computing Machinery. ISBN 9781450399906. doi: 10.1145/3570361.3592527. URL https://doi.org/10.1145/357...

    work page doi:10.1145/3570361.3592527 2023

  52. [52]

    Rf-3dgs: Wireless channel modeling with radio radiance field and 3d gaussian splatting, 2026

    Lihao Zhang, Haijian Sun, Samuel Berweger, Camillo Gentile, and Rose Qingyang Hu. Rf-3dgs: Wireless channel modeling with radio radiance field and 3d gaussian splatting, 2026. 13

    work page 2026

  53. [53]

    Newrf: a deep learning framework for wireless radiation field reconstruction and channel prediction, 2024

    Haofan Lu, Christopher Vattheuer, Baharan Mirzasoleiman, and Omid Abari. Newrf: a deep learning framework for wireless radiation field reconstruction and channel prediction, 2024

    work page 2024

  54. [54]

    Geo2SigMap: High-Fidelity RF Signal Mapping Using Geographic Databases

    Yiming Li, Zeyu Li, Zhihui Gao, and Tingjun Chen. Geo2SigMap: High-Fidelity RF Signal Mapping Using Geographic Databases. In 2024 IEEE International Symposium on Dynamic Spectrum Access Networks (DySPAN), pages 277–285, 2024. doi: 10.1109/DySPAN60163. 2024.10632773

    work page doi:10.1109/dyspan60163 2024

  55. [55]

    Recommendation ITU-R P.2040-3: Effects of building materials and structures on ra- diowave propagation above about 100 MHz

    ITU-R. Recommendation ITU-R P.2040-3: Effects of building materials and structures on ra- diowave propagation above about 100 MHz. Technical report, International Telecommunication Union, August 2023. URLhttps://www.itu.int/rec/R-REC-P.2040-3-202308-I/en

    work page 2040

  56. [56]

    No 3D” indicates whether the method operates without a 3D scene mesh at inference. “Per-path Delay

    Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks, 2018. URL https: //openreview.net/forum?id=H1bM1fZCW. 14 A Extended related work and comparison Deep learning approaches to channel modeling fall into three categories. We explore each in turn to giv...

    work page 2018