Learning to Twin Wireless Channels: Conditional Synthesis for UAV-to-Vehicle Communication

Accurate channel impulse response (CIR) data is critical for developing and optimizing unmanned aerial vehicle (UAV)-to-vehicle communication systems. However, acquiring sufficient measurements in the real world is often costly and time-consuming. This paper presents ChannelWGAN, a novel digital twin (DT)-enabled framework for high-fidelity CIR generation using a conditional generative adversarial network (cGAN). Our model uniquely conditions the generation process on key geometric parameters: pitch angle, azimuth angle, and slant range, to establish a physically meaningful mapping from the spatial domain to the channel response. To overcome the challenge of capturing complex temporal dependencies in CIR sequences, we integrate a Transformer architecture into the generator of a Wasserstein generative adversarial network with gradient penalty (WGAN-GP). This integration ensures stable training and superior modeling of long-range temporal correlations inherent in wireless multi-path propagation. The trained generator acts as an efficient channel twin, capable of synthesizing unlimited realistic CIR samples on demand for simulation purposes. Rigorous quantitative and qualitative evaluations in the baselines and the ablation study are conducted on a proprietary UAV-to-Vehicle dataset. The generated CIRs exhibit remarkable consistency with the real measurements in both visual patterns and key statistical properties. This work provides a robust and scalable data-driven solution for augmenting channel datasets, facilitating the development of future UAV communication systems.