Generative-AI Methods for Channel Impulse Response Generation

In this work, we propose methods for generating and manipulating channel impulse responses using normalizing flows. Using standardised, simplified, analytic models, when no perfect description of the channel is known, can lead to performance losses. We are able to show using simulations that our machine learning methods generate channel impulse responses with forced features. In addition to that, we show how disentanglement in the latent space of a normalizing flow can be used for the changing of certain features. We evaluate our methods using the maximum mean discrepancy.