Making UAV Wireless Models Explainable: Bridging Physics and AI with AERPAW

Gunes Karabulut-Kurt

Professor, Polytechnic Montreal

Gunes Karabulut-Kurt, Professor at Polytechnique Montréal, together with her students Kursat Tekbiyik and Antoine Lesage-Landry, is tackling a fundamental challenge in wireless communications for UAVs: balancing accuracy with interpretability. Modern deep learning models can achieve high prediction accuracy, but they often function as “black boxes,” offering little insight into the underlying physical behavior of wireless signals. In contrast, traditional physics-based models are transparent and explainable but too simplistic to capture the complexity of real-world UAV propagation environments.

To bridge this gap, the team developed an innovative approach based on Kolmogorov-Arnold Networks (KAN), which uses learnable functions along network edges to generate transparent, symbolic mathematical expressions. This enables the model to remain interpretable while still capturing complex patterns in the data. Importantly, their approach incorporates established physical laws—such as Free-Space Path Loss and Two-Ray propagation models—as adaptable biases, ensuring that the learning process remains grounded in electromagnetic theory rather than relying solely on data-driven approximations.

AERPAW played a critical role in enabling this research by providing rich, real-world datasets and experimental environments. The team leveraged the AERPAW UAV Air-to-Ground (A2G) dataset, which includes precise 3D positioning and multi-frequency signal measurements, allowing them to train and evaluate their models under realistic conditions. In addition, AERPAW’s outdoor testbed environments introduced diverse terrain features and naturally “noisy” data, which were essential for demonstrating the robustness and flexibility of their Physics-Informed KAN (PIKAN) approach compared to rigid theoretical models. Through this work, the team demonstrates a powerful new direction for wireless modeling—one that combines the interpretability of physics with the adaptability of machine learning—paving the way for more reliable, explainable, and scalable UAV communication systems in complex real-world environments.

This research is captured in their recent and forthcoming publications, including “PIKAN: Physics-Inspired Kolmogorov-Arnold Networks for Explainable UAV Channel Modelling,” presented at the IEEE Aerospace Conference in March 2026, and “Physics-Inspired Kolmogorov-Arnold Networks based UAV Channel Models,” planned for submission to IEEE Transactions on Wireless Communications in May 2026.