Using Data Science and Machine Learning to Predict the Failure Rate of Pin Diodes

GaN research is important to the development of next generation power conversion technology because its large Baliga figure of merit GaN gives it the potential to outperform SiC. Presently, commercialized GaN technology is based on lateral high electron mobility transistors (HEMTs); however, large areas would be required for high voltage applications thus there is motive to switch to vertical GaN. Current GaN substrate manufacturing technology is known to produce inconsistent quality with a varying concentration of defects. These defects are known to cause catastrophic device failure, thus there is motive to develop quick, non-destructive techniques to predict the quality of the wafer before undergoing the expensive fabrication process. This talk focuses on using data science and machine learning to predict the quality of the diodes. This work demonstrated that Raman spectroscopy can be used to detect high crystal stress points, which strongly correlates with an increased leakage current. Optical profilometry images can be used to detect defects that cause catastrophic failures; however, the presence of benign defects makes it difficult to predict the device quality using a simple algorithm. Thus, a machine learning algorithm relate optical profilometry images to device performance of subsequently processed device. The results showed that this algorithm was 91% accurate at predicting the forward bias behavior of devices, and it is possible that the devices it could accurately predict failures due to device processing errors. This work was supported by the Office of Naval Research.