Smart Input Space Sampling Combined with Kriging-Partial Least Square Regression for EMC Risk Analysis at PCB Level with Many Variables

This article describes an advanced methodology for training surrogate models (SMs) in order to study the sensitivity of many geometrical parameters of printed circuit boards (PCBs) for a quicker investigation of routing rules regarding potential electromagnetic compatibility (EMC) or signal integrity (SI) issues. This methodology consists in two steps. First, it retains only the sensitive parameters of the input space that mostly explain the several output observations. Second, it consists in training iteratively several competing SMs, selecting the most efficient one, regarding training time and/or accuracy. Here, the Kriging-partial least square Regression (K-PLS) is identified as a relevant method for the considered case study, which consists of a 17-parameter multiple outputs crosstalk scenario. An extreme value analysis is provided as an outcome for a possible adjustment of design rules.