The paper presents a new method for modeling the functionality and Signal Integrity (SI) performances of bus communications at a high level of abstraction. Using neural networks, we build a SystemC/SystemC-AMS virtual platform that combines functional modules, which represent the operative behavior of the system, and non-functional modules based on neural network approximation, which display the systems SI characteristics. Our method was demonstrated by modeling a Universal Serial Bus 3.0 (USB 3.0) system and was applied to the prediction of transient waveforms and eye diagrams. Compared to a HSPICE simulation, our method achieves excellent accuracy (mean absolute error of 1%) with a much shorter simulation time (×6000 speedup).