PHYSICS-GUIDED NEURAL RADIO MAPS FOR CLOSED-LOOP VEHICULAR NETWORK DIGITAL TWINS
Infocommunication technologies and electronic engineering · 2026-11 · 6 (2) · 1-17
DOI: 10.23939/ictee2026.02.001 ↗Kurz gefasst
Fachartikel in der Zeitschrift Infocommunication technologies and electronic engineering, 2026-10. Verfasst von Ivan Fesiuk und 1 weiteren Autor:innen. Die Arbeit nennt 28 Literaturangaben.
Abstract (Originalsprache)
Considering the importance of communications for driving safety, autonomous vehicles are evaluated today in co-simulation platforms that pair a photorealistic driving simulator with a discrete-event network simulator. However, the driving simulator models every building, whereas the network simulator typically does not, since its radio channel is a statistical function of distance alone, so that a warning transmitted across an open square and one transmitted through a city block are received with the same probability. In this paper, we propose a novel physics-guided neural radio map that serves as the wireless propagation model of the intelligent transportation system digital twin. The building geometry extracted from the driving world is used as an environmental constraint for the wireless network simulation in ns-3, which is further extended by a neural network surrogate trained with residual damping and a reciprocity constraint to learn the signal propagation behavior from sparse measurements. The proposed approach is based on training the model on a sparse virtual measurement experiment synthesized from the building geometry extracted by the running driving simulator and deploying it as a precomputed four-dimensional table, which enables real-time execution of the vehicle-to-everything (V2X) channels in the CARLA and ns-3 co-simulation. Simulation results for the CARLA Town10 scene coupled with ns-3 show that the proposed approach allows us to reach the accuracy of the strongest neural baseline with several times fewer links at every experiment size from 40 to 5000 links, even when the reference channel cannot be expressed by the physical model and the measurement experiment is realistically constrained. In addition, the learned field remains nearly reciprocal, with a reciprocity violation of only 0.92 dB RMS against 4.9–7.1 dB for the multilayer perceptron (MLP) baselines and 21 dB for the black-box model and reaches a coverage intersection over union (IoU) of 0.91 against 0.86 for the best MLP, while the deployed table answers a path-loss query independently of network size and model depth.
Zitieren
APA
Fesiuk, I., & Maksymyuk, T. (2026). PHYSICS-GUIDED NEURAL RADIO MAPS FOR CLOSED-LOOP VEHICULAR NETWORK DIGITAL TWINS. Infocommunication technologies and electronic engineering, 6(2), 1-17. https://doi.org/10.23939/ictee2026.02.001
Vancouver
Fesiuk I, Maksymyuk T. PHYSICS-GUIDED NEURAL RADIO MAPS FOR CLOSED-LOOP VEHICULAR NETWORK DIGITAL TWINS. Infocommunication technologies and electronic engineering. 2026;6(2):1-17. doi:10.23939/ictee2026.02.001
BibTeX
@article{Fesiuk2026,
author = {Fesiuk, Ivan and Maksymyuk, Taras},
title = {PHYSICS-GUIDED NEURAL RADIO MAPS FOR CLOSED-LOOP VEHICULAR NETWORK DIGITAL TWINS},
journal = {Infocommunication technologies and electronic engineering},
year = {2026},
volume = {6},
number = {2},
pages = {1--17},
doi = {10.23939/ictee2026.02.001}
}
EUVIMED