Publication: Materials characterization and aiaided defect classification of pad discoloration in semiconductor wafer fabrication
Loading...
Date
2025-09-16
Authors
_Izzuddin Iskandar, Norizan
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The discoloration of metal bond pads in semiconductor wafer fabrication is a constant problem because it can be caused by a number of things, from harmless native oxide formation and copper wrinkling to severe corrosion, embedded residues, or surface contamination that make wire bonding less reliable. Automated Optical Inspection (AOI) systems can see color changes, but they can't always tell the difference between low-risk and high-risk defects. This means that they might miss important defects or lose yield. This study gets around this problem by combining advanced material characterization with an AI-enhanced AOI classification framework to improve wafer yield without lowering reliability. We collected AOI images from separate diode and IGBT wafers and put them into seven groups based on the types of defects they had: Non-Visible, Pad Residue, Corrosion, Particle, Copper Wrinkling, Pad Deformation, and Scratches. We used optical microscopy, scanning electron microscopy (SEM) with energy-dispersive X-ray spectroscopy (EDX), and focused ion beam (FIB) cross-sectioning to find out how bad the defects were. Then we used Failure Mode and Effects Analysis (FMEA) to measure how bad they were. We used these FA-derived classifications to teach a ResNet-18 convolutional neural network (CNN) model on the carefully chosen AOI dataset. The trained AI model was 98.5% accurate overall, 94.8% accurate at selective inking, and had a 0.07% chance of an escapee. Grad-CAM visualizations showed that the network always focused on the right defect features, making it strong against changes in image brightness, contrast, and surface texture. The AI-powered classification system increased wafer yield from 65% to 85%, cut down on false positives, and kept strict detection of high-risk defects. This study shows that combining FA-based defect knowledge with AI-enhanced AOI makes it possible to tell the difference between cosmetic and critical pad discoloration in real time and on a large scale. This is a solution for semiconductor quality control that balances yield optimization with reliability assurance and sets the stage for more AI use in high-volume wafer manufacturing.