Wideband Kriging for Multiobjective Optimization of a High-Voltage EMI Filter

Employing multiobjective optimization (MOO) for grid-based 3-D electromagnetic simulations is computationally expensive. Using machine learning models as surrogates can mitigate this problem. However, for complex electromagnetic compatibility (EMC) systems, surrogates become inefficient since the required number of training samples leads to high model generation times. Furthermore, in EMC, the functions are often described in a frequency range. This demands a vectorial approach to the problem, which most methods cannot provide. In this work, we propose a Gaussian process model, called wideband Kriging, to generate a vector-valued model over frequency. We study, as an example, a high-voltage electromagnetic interference (EMI) filter's MOO over a design space of 17 parameters. The wideband Kriging needs only 450 training samples to achieve an accuracy of <1.5% for a frequency range from 1 kHz to 400 MHz. The model runs 800 times/minute, enabling its use for high dimensional MOOs. We find the best filter design concerning volume, current rating, electromagnetic compatibility (EMC) performance, and saturation.