Enhanced Defect Detection Using ResNet-Based Neural Networks After Wafer Bonding

Wafer bonding is key to realizing integration for smart systems. Using the wafer direct bonding technology, bonding between materials such as GaAs-to-Si, GaN-to-Si and SiC-to-Si along with Si-to-Si can be realized. Most of these bonding methods realize electrical connection with the help of copper interconnects and small defects at the bond interface would lead to the non-functionality of the devices, rendering them useless. Hence, quick remediation in analysing the bond defect with high accuracy and classification of such defects would decrease the time for evaluation and process optimization. In this paper, we demonstrate our initial work on wafer defect detection. The dataset is captured using an infrared (IR) camera which makes it quick to analyze the bond interface for any abnormality. A ResNet-based artificial intelligence model for panoptic segmentation is then adapted by training on this data for defect identification and classification. Our ResNet-based model can identify and make a distribution plot of the defects based on the class defined with up to 95% accuracy for large-size defects.