With the advent of 6G communication, Ray-Tracing technology has become a significant area of interest in computational graphics. In this context, radio base stations are expected to estimate wireless ray trace reflections in the surrounding environment accurately. To this end, recent advances in Omni-verse have inspired neural surrogates that model wireless electromagnetic propagation effects in physical environments. These surrogates provide a fast, differentiable, and continuous digital twin that enhances state-of-the-art ray-tracing technologies.
Given such development, the neural surrogate of conventional raytracer is still not efficient. Just like conventional raytracer, one still needs to shoot millions of rays to get tens of effective rays, wasting most of the computation effort.
Graph Neural Networks (GNNs) have proven successful in learning representations of complex systems. Recent research has also shown successful applications of GNNs in representing physical worlds with grids and geometric shapes. The aim of this project is to develop a GNN with geometrical features that can learn the ground truth of ray-tracing without the need for exhaustive computations.
A possible way is to structure the exhaustive ray emission from the base station into a graph structure and use GNN to filter out the irrelevant rays in the environment. This could possibly include an auto-encoder to simplify the ray computation based on its understanding of the physical environment.