A Machine Learning based Channel Modeling for High-speed Serial Link
Xiao Li, Qingsheng Hu · 2020
In high-speed signal transmission, a large amount of channel data is needed to optimize the serial link and improve the signal integrity. To overcome some drawbacks of traditional channel modeling, such as inconvenient and long simulation time, a machine learning based channel modeling is presented, which can improve the efficiency but also ensure the accuracy. By using random forest regression (RFR) algorithm and selecting several feature variables that affect channel characteristics as its inputs, an output in terms of S-parameter at each frequency point of the channel can be obtained. Furthermore, an improved modeling method, in which accurate model instead of rough one is employed for some interested frequency range, is proposed to make tradeoff between the accuracy and efficiency. Experiments results illustrate that our RFR based method can speed up the modeling process significantly. Meanwhile, the accuracy is ensured by combing the rough and accurate model effectively.