Deep Learning-Accelerated Ray Tracing

This work was supported by an NSERC reserach grant. It was performed under the supervision of Prof. Costas Sarris as part of a summer research project.

Ray-tracing is a computationally expensive process, especially in complex geometric environments where many reflections, transmissions, and diffractive effects (collectively referred to as ray-interactions) must be accounted for. In this project, we limited the number of ray-interactions in our ray-tracer runs to produce low-fidelity maps of recieved signal strength (RSS) from a wireless access point (WAP). The RSS maps were then passed to a physics-informed generative adversarial network (GAN) to improve their fidelity and accuracy. Overall, we saw a reduction in runtime of 85% with a 4% relative error against ground-truth.

Read the paper here.