Modeling the Impact of Phase Noise in THz Communications Through Machine Learning: An Approach for Efficient Waveform Design

Phase noise (PN) poses a major challenge in terahertz (THz) communications by degrading bit error rate (BER) and system capacity. This study utilizes machine learning (ML) to evaluate the effect of PN in a THz communication system. An in-house developed custom PN simulator is utilized to generate training and test data with a variety of waveform design parameters, with BPSK modulation. A highly accurate regressor is developed and paired with Differential Evolution (DE) optimization, whereby a framework to predict the system performance as well as identify the most important features in waveform design is setup. It is shown that the framework can account for the contradictory nature between conventional additive white Gaussian noise (AWGN) and PN, and provide critical information needed for optimal waveform design in the THz regime. This work may open the door for future utilization of ML-based regressors in THz physical and MAC layer design.