Reverse characterization prediction of diamond Schottky barrier power devices using machine learning: Predicting breakdown voltage and Baliga figure of merit

Abstract

Diamond power devices showed great potential for high-performance and low-loss applications at high frequencies. Schottky power diodes on the diamond have markedly higher threshold reverse operation limits than other power devices, yet they are still far from the ideal limit. Therefore, utilizing the device capabilities across optimizing device parameters is severely time- and resource-intensive considering traditional trial-and-error experimental methodology. On the other hand, data-driven techniques can accelerate device optimization and reduce resource consumption. This work demonstrates machine learning models that predict breakdown voltage in addition to Baliga's power figure of merit (BFOM) for the first time. An experimentally driven dataset is constructed because of its reliability, precision, and potential for generalization. This dataset contains current-voltage characteristics raw data for more than 400 devices fabricated on heteroepitaxial diamond. Four machine learning algorithms are explored and evaluated based on 5-fold cross-validation to prevent overfitting. The extra-tree model shows the highest performance that accurately predicts the breakdown voltage, and the BFOM with R2 equal to 0.924 and 0.911, respectively. Additionally, it showed robust prediction accuracy on unseen data with an 8 % average relative error of 16 devices. Achieving this high accuracy indicates the potential of these models for generalization on various datasets.