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

Accurate frequency response estimation is crucial in various engineering applications, including electrical circuits especially compact power electronic circuits. In EMC, numerical simulations are often computationally expensive, especially if they need to be run multiple times for optimization. Trained machine learning models offer a suitable solution. With such a model, millions of simulations can be run per day, enabling to perform multi-objective optimization even for complex physical models. However, these models are particularly accurate only at "well-behaved" areas of the transfer function but not at their resonances. This paper investigates a methodology which incorporates the systems physics to better resolve these resonances. Instead of predicting the frequency response on individual frequency samples this work aims to predict the coefficients of its complex-valued transfer function. The results show that the physics-informed approach can resolve the resonances of the transfer function better, but the accuracy of the predicted coefficients must be very high.