Anomaly Detection in Semiconductor Process Validation Using Unsupervised Learning and Generative Models

Abstract

Semiconductor manufacturing is characterized by highly complex, multistage processes that demand stringent validation to maintain product quality and yield. As traditional supervised approaches require extensive labeled datasets, there is an increasing interest in leveraging unsupervised learning and generative models for anomaly detection. This paper explores the integration of these advanced methods into semiconductor process validation circa 2020, addressing the challenges of high-dimensional data, subtle fault patterns, and label scarcity. By employing unsupervised learning techniques, such as clustering and autoencoders, alongside generative models like GANs and VAEs, the study demonstrates notable improvements in early fault detection rates. Methodologies involve the use of historical process data without explicit fault labeling, enhancing model adaptability to unseen anomalies. Our findings underline the potential of these approaches to achieve higher sensitivity while reducing false alarms compared to traditional methods. This research contributes to advancing the field toward more autonomous, reliable validation frameworks.