Frequency Response Estimation via Multi-Coefficient Surrogate Models of Rational Complex Functions

Accurate frequency response estimation is crucialin various engineering applications, including electrical circuitsespecially compact power electronic circuits. In EMC, numericalsimulations are often computationally expensive, especially ifthey need to be run multiple times for optimization. Trainedmachine learning models offer a suitable solution. With such amodel, millions of simulations can be run per day, enabling toperform multi-objective optimization even for complex physicalmodels. However, these models are particularly accurate onlyat "well-behaved" areas of the transfer function but notat their resonances. This paper investigates a methodologywhich incorporates the systems physics to better resolve theseresonances. Instead of predicting the frequency response onindividual frequency samples this work aims to predict thecoefficients of its complex-valued transfer function. The resultsshow that the physics-informed approach can resolve theresonances of the transfer function better, but the accuracy ofthe predicted coefficients must be very high.