ReVeal: A Physics-Informed Neural Network for High-Fidelity Radio Environment Mapping

Mukaram Shahid, Kunal Das, Hadia Ushaq, Hongwei Zhang, Jiming Song, Daji Qiao, Sarath Babu, Yong Guan, Zhengyuan Zhu, Arsalan AhmadDySPAN 2025 · DOI

TL;DR

This paper introduces ReVeal, a novel Physics-Informed Neural Network (PINN) that accurately maps radio environments using sparse RF sensor data. By integrating a derived partial differential equation (PDE) for path loss into the neural network's loss function, ReVeal achieves significantly higher accuracy than existing methods with minimal data and computational cost, paving the way for more efficient spectrum management.

What Problem Does This Paper Address?

Accurately mapping the radio environment, which includes identifying wireless signal strength across different frequencies and locations, is crucial for efficient spectrum sharing. Current methods, like statistical or deterministic models, struggle with accuracy and generalization due to factors like shadowing, interference, and fading. Deterministic models, such as ray tracing, are computationally expensive and require highly detailed environmental information that is often unavailable. Statistical models, while less computationally demanding, lack precision and fail to capture site-specific variations. Existing data-driven approaches like Kriging and deep learning also have limitations, including assumptions of spatial uniformity or the need for vast amounts of training data. This gap highlights the need for a model that is accurate, efficient, generalizable, and requires minimal data and prior knowledge of transmitter parameters or environmental details.

Methods

The core innovation of ReVeal lies in its use of a Physics-Informed Neural Network (PINN) architecture. The authors first derive a second-order partial differential equation (PDE) that models the Received Signal Strength Indicator (RSSI) based on a statistical path loss model, incorporating shadowing effects. This PDE is then integrated into the neural network's loss function as a "physics loss." The overall loss function is a weighted combination of this physics loss and a data-driven loss (mean squared error between predicted and observed RSSI). This approach allows the neural network to learn from sparse real-world measurements while simultaneously enforcing physical consistency through the PDE. The architecture uses a standard feed-forward neural network with dropout layers for regularization. Hyperparameter tuning is performed using the Optuna library.

Datasets

The paper evaluates ReVeal using real-world measurement data collected from the ARA (Advanced Research on Wireless) testbed located around Iowa State University. This testbed covers an area of approximately 514 square kilometers, encompassing diverse rural and suburban terrains. Data was collected using Skylark mMIMO systems deployed at the Wilson Hall base station, with other base stations turned off during data collection to isolate the signal source. The sampling strategy employed the Local Pivotal Method (LPM) to select informative spatial points.

Key Insights and Ideas

  • Physics-informed learning for radio mapping: Integrating physical laws (via a PDE) directly into the neural network's training process is a powerful way to improve accuracy and generalization, especially with sparse data.
  • PDE captures shadowing effectively: The derived second-order PDE, unlike simpler empirical path loss models, is shown to be crucial for accurately capturing the complex spatial variations in signal strength caused by shadowing.
  • Data efficiency is paramount: ReVeal demonstrates that high accuracy can be achieved with a remarkably small number of training samples (e.g., 30 points over 514 km²), making it practical for real-world deployment.
  • Balancing data and physics: The parameter $\lambda$ in the loss function is critical for balancing the influence of data-driven observations and physics-based constraints, highlighting the importance of fine-tuning this trade-off.

Key Results

ReVeal achieved a Root Mean Square Error (RMSE) of 1.95 dB, which is an order of magnitude lower than existing methods like 3GPP TR 38.901 (17.25 dB RMSE) and ITU-R IMT-2020 (11.13 dB RMSE). It also significantly outperformed Kriging (12.02 dB RMSE) and a standard Fully Connected Neural Network (FCNN) (10.59 dB RMSE). Computationally, ReVeal is highly efficient, requiring only 8.9 seconds for training and prediction, comparable to statistical models and much faster than ray-tracing or FCNN. The model demonstrated robust performance across varying terrains and was able to capture RSSI variations effectively, even in complex scenarios.

Limitations and Future Work

The authors acknowledge that temporal variations caused by fading are beyond the scope of this study and are treated as a part of future work. They suggest that future research could focus on applying ReVeal to dynamic spectrum sharing mechanisms and incorporating temporal dynamics to predict spectrum usage trends. The study also highlights the importance of fine-tuning the $\lambda$ parameter to effectively balance data-driven and physics-driven loss terms for optimal prediction.

Summary updated

BibTeX
@inproceedings{shahid2025reveal,
  author = {Mukaram Shahid and Kunal Das and Hadia Ushaq and Hongwei Zhang and Jiming Song and Daji Qiao and Sarath Babu and Yong Guan and Zhengyuan Zhu and Arsalan Ahmad},
  title = {{ReVeal: A Physics-Informed Neural Network for High-Fidelity Radio Environment Mapping}},
  booktitle = {DySPAN},
  year = {2025},
  doi = {10.1109/dyspan64764.2025.11115911},
}
Figure