Wafer defect localization and classification using deep learning techniques

Wafer defect localization and classification is a critical task in semiconductor manufacturing, where identifying and categorizing defects on wafers is essential for maintaining product quality. This study presents a deep learning-based approach using YOLOv2 (You Only Look Once version 2) for the detection and classification of wafer defects. A comprehensive dataset sourced from Google was used, with ground truth labels created through the Ground Truth Labeller App, ensuring precise defect annotations. The dataset, stored with accurate defect labels, includes various defect types, such as Edge Local, Edge Ring, Donut, Center, Local, Scratch, Random, Near Full, and None. The YOLOv2 object detection model, known for its high efficiency and accuracy in real-time object detection, was implemented with specific layers tailored for wafer defect localization. The model was trained using CNN layers optimized for this task, employing well-defined training options to ensure high classification accuracy. Each defect type was effectively classified, contributing to the detection system's ability to localize and categorize defects with a high degree of precision. The final model demonstrated promising performance with significant accuracy, offering a robust solution for wafer defect detection in semiconductor production, potentially reducing errors, enhancing efficiency, and ensuring high-quality manufacturing. This method provides an automated, reliable, and scalable solution for real-time wafer defect inspection, facilitating further advancements in semiconductor industry automation.

Keywords: Dataset, Image Processing Techniques, Deep Learning, YoloV2 Detection and Convolution Neural Network.