Comparison of Machine Learning Techniques for Analysing Wide Bandgap Switching Behavior and EMI in Half-Bridge Structures

Wide bandgap semiconductors allow the design of highly efficient and power-dense converters due to their fast switching, high frequency operation. This, however, gives rise to significant electromagnetic interference (EMI) that degrades the power system. Accurate prediction of switching behavior and EMI generation is challenging due to their dependence on device and circuit parasitics. Simulations can address this; however, design optimization that requires the consideration of numerous design variations makes this a time-intense process. Machine learning can achieve this more efficiently, however there are many techniques available that vary in their suitability for such a task. This paper presents a comparison of various machine learning algorithms used to predict the switching transition waveforms and harmonic spectrum of a SiC-based half bridge structure. Results show the Multi-Head Attention Artificial Neural Network emerging as a suitable method due to its consistently high predictive accuracy across a range of seen and unseen MOSFET datasets.