Modeling Millimeter Wave Channels with Generative Adversarial Networks

Modern cellular systems increasingly rely on concurrent communication across several discontinuous bands due to broader bandwidth and macro-diversity. Multifrequency communication is crucial in millimeter wave (mmWave) and Terahertz (THz) frequencies, frequently paired with lower frequencies for resilience. Statistical models capable of representing the combined distribution of channel routes over many frequencies are needed to assess these systems. This study presents a broad neural network-based training approach for multifrequency double-directional statistical channel models. The suggested method involves representing every channel as a multi-clustered set and training a generative adversarial network (GAN) to generate random multi-cluster profiles. The resulting cluster data consists of vectors distributed at various frequencies with random received powers, angles, and delays. Urban micro-cellular connections at 28 and 140 GHz are modeled using ray tracing data to demonstrate the methodology. The model is readily adaptable for multi-frequency link or network layer simulation. Studies show that the model may capture statistical correlations between frequencies, and this technique involves minimal statistical assumptions.