Machine learning approach for sorting SiC MOSFET devices for paralleling

Abstract

This paper presents the development of a machine learning model for sorting SiC MOSFET devices for paralleling. A multivariate linear regression model is developed and trained with device parameter data (as input) and current imbalance data (as label). Each of the training devices is successively paralleled with one reference device to generate the current imbalance (label) data. Devices with close values of current imbalance when individually paralleled with the reference device are expected to have relatively balanced current sharing when paralleled among themselves. This model is trained with 40 devices and tested with 20 devices. The model shows accuracies of 87.93% and 97.48% in predicting transient and steady state current imbalances, respectively. These accuracy values are obtained by comparing the model’s prediction for the 20 testing devices with the actual current imbalance values when these devices are paralleled with a reference device. Based on current imbalance predictions made by these models, the devices are grouped into classes for paralleling applications. The selection of devices for paralleling based on predictions from these models result in a satisfactorily balanced current distribution. The model’s performance at higher temperatures is also satisfactory.