Self-Supervised Learning for Automated Hardware Defect Detection in Semiconductor Manufacturing

Abstract

The increasing complexity of semiconductor manufacturing processes necessitates advanced defect detection mechanisms to ensure high yield and quality. Traditional defect detection methods, often reliant on rule-based or supervised learning models, face limitations in handling diverse defect patterns and require extensive labeled datasets. To address these challenges, this study proposes a Self-Supervised Learning (SSL)-based automated hardware defect detection framework that leverages contrastive learning and masked autoencoders to identify anomalies in semiconductor wafers and circuits. The proposed approach learns rich feature representations from unlabeled data, reducing dependency on manual annotations while enhancing defect localization accuracy. Extensive experiments on real-world semiconductor defect datasets demonstrate that the SSL-based model achieves higher defect detection accuracy (+12% improvement over supervised CNN models), better generalization, and lower false positive rates. Additionally, the framework significantly reduces the need for labeled data, making it highly suitable for scalable industrial applications. This research highlights the potential of selfsupervised AI techniques in semiconductor defect inspection, paving the way for more efficient, cost-effective, and automated hardware quality assurance systems.