Generalizable Neural Ray Tracing Towards Physics-Informed Intelligent Channel Modeling

Neural ray tracing (RT) has emerged as a powerful paradigm for intelligent channel modeling by combining physical propagation principles with neural networks. However, existing approaches often suffer from strong spatial dependence that hinders generalization and exhibit limited ability to capture the fundamental electromagnetic behavior laws. In this paper, we propose GeNeRT, a Generalizable Neural RT framework that addresses these issues through two key innovations. The first is an outgoing ray prediction network that takes as input the relative geometric features between rays and surfaces, thereby decoupling spatial dependency. Second, a Fresnel-inspired dual-branch architecture, coupled with a two-stage training strategy, is employed to accurately model the ray-surface interactions. Experiment results demonstrate that GeNeRT effectively captures universal polarization behaviors and significantly outperforms prior neural RT methods across diverse test conditions, achieving both high modeling accuracy and strong generalization capability in outdoor scenarios.