RadioLunaDiff
Estimation of wireless network signal strength in lunar terrain
This work was done in the UW Sensor Systems Lab in collaboration with several other talented researchers, namely Paolo Torrado. As an undergraduate, I built a data generation pipeline, using my own custom terrain generation algorithm and Sionna-RT. I also developed and implemented the NN prediction architecture. In September of 2025 we submitted a paper to ICASSP.
This project introduces a novel physics-informed deep learning architecture for predicting radio maps (wireless signal strength) over complex lunar terrain. This work supports future communication-aware mission planning for NASA’s proposed LunaNet framework.
Abstract
In this paper, we propose a novel physics-informed deep learning architecture for predicting radio maps over lunar terrain. Our approach integrates a physics-based lunar terrain generator, which produces realistic topography informed by publicly available NASA data, with a ray-tracing engine to create a high-fidelity dataset of radio propagation scenarios. Building on this dataset, we introduce a triplet-UNet architecture, consisting of two standard UNets and a diffusion network, to model complex propagation effects. Experimental results demonstrate that our method outperforms existing deep learning approaches on our terrain dataset across various metrics.
System Architecture
Our approach uses a pipeline of three cascaded models—two UNets followed by a diffusion network—to predict a radio map. The first UNet estimates the binary square wave number map ($k^2$). The second UNet produces an initial radio map estimate. Finally, a Denoising Diffusion Probabilistic Model (DDPM) refines the initial map by predicting the residual.
Example Inputs & Outputs
The model takes several inputs, including the terrain height map, a high-pass filtered version of the terrain, and a one-hot map of the transmitter’s location.
Below are the intermediate outputs from the model (at 415 MHz), showing the predicted $k^2$ map, the initial coarse radio map, and the predicted residual used for refinement.
Finally, here is a comparison of the model’s final refined prediction against the ground truth (from the Sionna ray-tracer) for both 415 MHz and 5.8 GHz frequencies.
