From Wafers to Bits and Back again: Using Deep Learning to Accelerate the Development and Characterization of SiC

Abstract

A non-destructive, fast and accurate extended defect counting method on large diameter SiC wafers is presented. Photoluminescence (PL) signals from extended defects on 4H-SiC substrates were correlated to the specific etch features of Basal Plane Dislocations (BPDs), Threading Screw Dislocations (TSDs), and Threading Edge Dislocations (TED). For our non-destructive technique (NDT), automated defect detection was developed using modern deep convolutional neural networks (DCNN). To train a robust network, we used our large volume data set from our selective etch method of 4H-SiC substrates, already established based on definitive correlations to Synchrotron X-Ray Topography (SXRT) [1]. The defect locations, classifications and counts determined by our DCNN correlate with the subsequently etch-delineated features and counts. Once our network is sufficiently trained we will no longer need destructive methods to characterize extended defects in 4H-SiC substrates.