Machine Learning Methods for Elaborating the Feasible Region for Signal Integrity Analysis in Differential Pair PCB Structures

High-speed printed circuit board (PCB) design has recently evolved more and more towards differential signaling, as this has certain advantages over single-ended signal transmission in terms of noise immunity, crosstalk susceptibility and ground isolation. Even though all of these enhancements contribute to a robust signal integrity (SI), differential signal transmission raises further questions and problems especially regarding skew and timing margins. This requires to carefully route the signals symmetrically often resulting in a complex and time-consuming task for the PCB designer. Under these conditions, machine learning (ML) methods have shown to be effective for prediction of the relevant SI performance metrics and can thereby assist and simplify the PCB design process. In this paper, neural network (NN) regression methods are utilized to predict the signal behavior in form of measurements of the resulting eye diagrams, skew and timing margins, common-mode voltages as well as differential time domain reflectometry (TDR) responses from the given design parameters. Based on the trained NN models a genetic algorithm (GA) is utilized for optimization especially with respect to the feasible region of the electrical parameters, i.e. the design constraints that satisfy the specified thresholds for eye-width, eye-height, intra-pair skew and common-mode voltage.