These slides were presented at the NAFEMS World Congress 2025, held in Salzburg, Austria from May 19–22, 2025.

Abstract

Electromagnetic theory has become indispensable for electromagnetic compatibility (EMC) optimization of electronic systems. For example, the road approval of electric vehicles is subject to strict electromagnetic emission and immunity limits. Computer simulation helps to understand the system complexity, the subtlety of the electromagnetic coupling paths, and to optimize electronic systems with respect to weight and cost. Using deterministic methods such as network or electromagnetic simulations, we can study realistic models of electronic systems, e.g. traction inverters of electric vehicles with simulations that have a wall-clock time of anything between minutes a several hours per run. This limits the number of samples that we can conveniently study within the industrial design cycle to a range from dozens up to several thousand at max. We show that is it possible to train machine learning models from realistic models of electronic systems which run with wall clock time in the order of milliseconds, independent of the underlying physical model being a circuit or an electromagnetic simulation. The gain in computational speed is at least a factor of 1000, enabling us to run millions of simulations a day. This raises the potential of simulation to a qualitatively new level, the implications of which are not yet fully understood. Using realistic models of electronic systems which are trained over up to 17-dimensional parameter spaces, we show how to solve multi-objective optimization problems of complex EMC scenarios. We discuss the obtained optimal solutions and the implication of these results on the EMC design of electronic systems. We also give an outline of the further potential that trained machine learning models in EMC may offer. In particular we talk about the benefit of uncertainty quantification and its implication of risk analysis in the design for electromagnetic compatibility. Further, the reverse problem, Bayesian inference appears to improve the notoriously bad prediction quality of EMC models. With many parameters not exactly known and hard to acquire, Bayesian inference allows to rigorously estimate the most likely set of model parameters to match to available measurements. We end the talk with a number of open challenges that wait to be solved.