A black-box approach for generating surrogate data for an amorphous-core inductor working up to magnetic saturation

This paper presents a black-box method for generating surrogate data from measurements taken on inductors working up to magnetic saturation. The proposed method is based on a sequential neural network architecture, optimized through the Python Tensorflow framework and the Keras API, which accurately predicts the inductor flux dynamics under varying operating conditions. The generated data can be useful to fit existing circuit models to a limited set of physical measurements (easy-to-measure quantities, i.e., inductor voltage and current), complemented by the obtained surrogate data. The proposed black-box model performs well in predicting flux across different frequencies and amplitudes. This research is a first proof of concept (we focus on zero-bias sinusoidal inputs and amorphous-core inductors at fixed temperature); anyway, it highlights the potential of combining deep learning with robust optimization and pre-processing techniques to improve predictive accuracy in circuit models of inductors working up to magnetic saturation. These models can be used for simulating and designing high-power-density switch-mode power supplies.