Train-to-Ground Wireless Channel Characterization Based on Physics-Inspired Generative Network

With the rapid advancement of intelligent high-speed rail (HSR), train-to-ground (T2G) wireless communication systems increasingly demand high reliability and speed. Accurate T2G channel modeling is crucial, particularly in AI-optimized and AI-native networks. To address this, we propose a physics-inspired generative neural network for wireless channel characterization, where the location information of users is used to predict the key channel parameters such as multi-path gain, delay, K-factor and Doppler shift. Unlike conventional methods requiring extensive channel data, our approach formulates prediction as a generative task, and the simulated propagation graph (PG) data is used for training. The architecture of the proposed network aligns with the physical channel impulse response (CIR) formulation, leveraging PG for scattering environment modeling and Sinusoidal Representation Networks (SIREN) to capture dynamic electromagnetic wave phase characteristics via sine activations. Experimental results demonstrate the effectiveness of the proposed model in accurately characterizing key channel features in high-speed mobile communication.