From Waves to Graphs: A Ray-Tracing-Inspired Neural Radio Propagation Model

Paul Almasan 2 , Stefanos Bakirtzis 3 , José Suárez-Varela 2 , Andra Lutu 2 2 Telefónica Research, 3 University of Cambridge · 2026

Read From Waves to Graphs: A Ray-Tracing-Inspired Neural Radio Propagation Model by Paul Almasan 2 , Stefanos Bakirtzis 3 , José Suárez-Varela 2 , Andra Lutu 2 2 Telefónica Research, 3 University of Cambridge (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-30525.txt.

Introduction

Paul Almasan and three co-authors present GRAPHWAVE, a neural graph-driven signal solver hinging on ray tracing rules. The model uses a digital signal area to build a point cloud and extract a matching graph of the radio area. By applying neural message passing over the matching graph, it infers well radio-related values, e.g., caught signal strength, in a 3D area. They showcase GRAPHWAVE as a radio area digital twin, showing it can learn from fake and real-world data with fast run times.

Catalog record

CREATORPaul Almasan 2 , Stefanos Bakirtzis 3 , José Suárez-Varela 2 , Andra Lutu 2 2 Telefónica Research, 3 University of Cambridge
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE28776
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.30525

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

From Waves to Graphs: A Ray-Tracing-Inspired Neural Radio Propagation Model Paul Almasan 2 , Stefanos Bakirtzis 3 , José Suárez-Varela 2 , Andra Lutu 2 2 Telefónica Research, 3 University of Cambridge Abstract Artificial intelligence-driven radio propagation models provide agile and robust solutions for mobile network operators in their effort to ensure the optimal performance of the wireless ecosystem and support its efficient expansion. In this paper, we introduce ural graph-driven propagation solver hinging on the governing principles of ray tracing. The proposed model leverages a digitized version of the propagation environment to build a point cloud and extract an equivalent graph representation of the radio environment. By applying neural message passing over the equivalent graph, it allows the model to accurately infer radio-related quantities, e.g. , received signal strength, in a three-dimensional environment. We showcase the use of adio environment digital twin and we demonstrate that the model can learn from synthetic and real-world data while achieving low inference times. I Introduction Artificial intelligence (AI) has transformed various aspects of the wireless ecosystem along with the methods and tools used to underpin its effective operation [ 11 ] . In particular, AI has enabled the emergence of radio access network (RAN) digital twins (DT) that provide predictive and data-driven representations of the propagation conditions experienced by the network. Notably, a drastic shift has been observed in radio propagation modeling, moving from conventional empirical or deterministic models to AI-powered solutions that exhibit enhanced accuracy, computational efficiency, and lower cost [ 10 ] . Such models yield an invaluable tool for mobile network operators (MNOs) to support predictive control and optimization of the RAN, enabling critical applications such as network planning, coverage optimization, interference management, beamforming, and localization [ 3 ] . To mold AI-powered radio propagation tools, one needs a model that accurately translates data from a radio propagation scene—with some transmitting devices deployed—into some radio quantities of interest (QoI), typically the received signal strength (RSS) or the path loss (PL). To this end, it is crucial to properly co-design a data representation that captures the fundamental factors influencing radio propagation in the scene, along with an AI model that can effectively learn and abstract patterns from these representations. Early works in this direction use tabular data along with multi-layer perceptrons (MLPs) [ 15 , 1 ] . Later works use two-dimensional (2D) image-like visualizations of the propagation environment and other physics-based input features, which can be combined with convolutional neural networks [ 19 , 12 , 2 ] or transformer-based encoding-decoding architectures [ 17 ] to exploit spatial correlations and infer the RSS. However, transforming the 3D propagation space into a simplified 2D scene inherently curbs the functionality and fidelity of radio propagation tools. First, 2D-based models constrain the prediction to a single receiving plane ( e.g. , at a predefined height). This renders them unsuitable for predicting radio propagation in the 3D space. Second, 3D geometries like buildings or cluttered data are reduced to pixels, hence losing structural details that can produce inaccuracies in the radio propagation modeling. Alternatively, more robust and realistic representations can be achieved via physics-driven approaches that consider the principles of radio propagation mechanisms. A well-established method is ray-tracing [ 18 ] , which simplifies the intricate behavior of wave propagation by assuming that electromagnetic (EM) energy flows along geometrically defined ray paths. These paths are created by the consecutive interaction of a ray/wave emitted from a transmitter with objects found within the propagation environment. Specifically, when a ray impinges onto an interface between two media, it can be partially reflected, refracted, diffracted, or scattered, and there exist laws and approximations to compute the strength and phase of each component [ 7 ] . In this paper, we present framework inspired by the principles of ray tracing that estimates wireless signal-related QoIs in 3D radio environments. The proposed framework directly integrates the radio scene and the ray paths into a unified graph representation, forming the basis of RAN DTs. To do that, we initially build a point cloud representation that captures all objects, transmitters, and receivers in the radio propagation scene. Then, these elements are used to construct a graph where point cloud elements are mapped to nodes in the graph, and edges interconnect all the scene elements by simulating rays (see schematic representation in Figure 1 ). Particularly, our work yields the following contributions: Fig. 1: System architecture overview of the proposed method, using various data modalities to construct a DPE, extract an equivalent graph representation, and apply a neural transformation on it to infer RSS. • We demonstrate how the digital representation of the propagation environment can be converted into a set of nodes, forming the foundations of our graph-driven propagation model. By establishing proper connections between the graph nodes, we instate a representation of the input data that innately resembles the potential paths EM waves can traverse. Unlike the widely adopted simplified 2D visualization [ 19 , 12 , 2 , 5 , 6 ] , our model favors from its physics-inspired architecture. • We design a set of custom node and edge features that enable our solver to learn to emulate radio propagation through neural message exchange. • We showcase the potential of arn from simulated and real-world measurement data, proving its reliability in a realistic operational setting using data from a production-grade network. II GraphWave Ideation and Problem Formulation res to establish an equivalent graph-driven representation of the propagation environment and the potential paths that EM waves may traverse. To achieve that, it is necessary to convert the continuous 3D physical world into a discretized digital equivalent space, referred to as the digital propagation environment (DPE). Specifically, the DPE includes three main components: ( i ) a set of physical objects, such as buildings, cars and trees, that interact with the EM waves, ( ii ) a transmitting antenna, T_{x} , and ( iii ) a set of receivers R . Each physical object is represented by a collection of triangular meshes in the 3D space, parameterized by its vertex coordinates (see Section III-A ). In Figure 1 we show a general overview of our entire framework, depicting how we build the DPE and extract a graph representation (see Section III-A ). Each point in the cloud accounts for a node, v , in a graph, G , that resembles the DPE. By properly arranging the connections between the graph nodes, we can create virtual trajectories that reenact the paths through which EM energy flows. Further, each node and edge connection between them is endowed with physics-based and network-related features (see Section III-B ). gamates the graph-driven representation with a neural transformation to infer a QoI value, \hat{y}_{r} , at each receiving point r\in R . This neural transformation is implemented through the exchange of learnable messages between the graph nodes, and corresponds to identifying a mapping function, f , that transforms a graph G into \hat{y}_{r} [ 8 ] . Notably, bits a schematic equivalence with ray-tracing, i.e. , graph paths correspond to potential ray trajectories, yet its neural nature allows our model to assimilate knowledge and learn from real-world data. Formally, the operation of fined as: \hat{y}_{r}=f(G;\theta),

Read the complete text