Adaptive Artificial Neural Networks for Power Loss Prediction in SiC MOSFETs

Silicon carbide (SiC) MOSFETs have become increasingly central in the development of high-efficiency power applications as a result of their unique properties. However, accurate modeling of power losses is a crucial aspect to fully benefit from the potential of $S$iC MOSFETs. Power loss estimation for devices with different chip sizes is a key research field that accelerates the pre-design of new products and evaluates their performance across various applications without the need for extensive physical testing. In parallel to the establishment of SiC power devices, machine learning (ML) techniques, as a branch of artificial intelligence (AI), offer promising advancements in optimizing the design and performance of modern electronics. This paper investigates the application of selected machine learning methods, specifically standard artificial neural networks (ANNs) and adaptive artificial neural networks (a-ANNs), to estimate conduction and switching losses in SiC MOSFETs. Experimental measurements at different electrical and thermal conditions are the input dataset to train and test the developed algorithms. The performance of traditional polynomial regression models is compared with ANNs and a-ANNs. The prediction results demonstrate that machine learning-based models outperform traditional methods, especially when only part of the dataset is provided in the training phase. The a-ANNs are the most accurate and flexible modeling approach and excellently perform when predicting the loss characteristics for thermal and electrical conditions never encountered in the training dataset. More in detail, a-ANNs predict the target variables with low percentage errors in the 5% range. These findings highlight the potential of machine learning in improving the estimation accuracy of power losses in $S$iC MOSFETs. Similar algorithms can benefit additional fields in the power semiconductor world.