This document describes a new methodology for building surrogate model dedicated to EMC analysis of printed circuit boards. Stochastic approaches have recently showed their interests for EMC phenomena simulations because of their ability to replace costly computations. The proposed method is composed of two process stages. The first one is a pre-processing step which consists in filtering inputs that are not significant regarding output variations. The second one consists in an iterative learning technique of the surrogate model avoiding the requirement for prior determination of the sample size. The method is tested from a representative scenario of an EMC problem. The gain in computation time offered by the method makes it possible to build surrogate models more efficiently.

This article presents a methodology using machine learning techniques for defining printed circuit board (PCB) design rules in order to reduce signal integrity (SI) or electromagnetic interference (EMI) issues. The scenario illustrating the situation for which these rules must be defined is modelled with a 3D EM solver available on the market and simulations are run with varying parameters in order to obtain a representative sample of the design space. This data set is then used to train a surrogate model (i.e. a metamodel) of the scenario based on kriging algorithm. Using this surrogate model, more than ten thousands of simulations are computed in a decent time. The surrogate model estimations allow to estimate the sensitivity of the varying parameters with respect to some specifications (crosstalk level and insertion loss). Finally, an analysis of output values for which some requirements (crosstalk level, insertion) loss are not fulfilled provide some insights about possible adjustment of guidelines in terms of parameter ranges. Finally, a practical design example is given to illustrate the methodology.