Data Representation and Preprocessing Effects on S-Parameter Modeling of High-Speed Channels Using Machine Learning

In this paper, the effects of data representations and preprocessing on machine learningbased S-parameter modeling of high-speed channels are investigated. Using a transformer network as a base model, two S-parameter representations in real/imaginary and magnitude/phase are compared and studied. Considering S-parameter data distributions, various preprocessing techniques including MinMax normalization, standardization, robust scaling, power transformation, and quantile transformation are compared and analyzed to improve accuracy. Moreover, the accuracy results are compared depending on the electrical length of target channels.