Equipment Anomaly Detection for Semiconductor Manufacturing by Exploiting Unsupervised Learning from Sensory Data.

Area of Science:

• Semiconductor Manufacturing
• Machine Learning
• Process Control

Background:

• In-line anomaly detection (AD) is crucial for semiconductor equipment maintenance and identifying potential line yield issues.
• Equipment sensory data (ESD) is vital for proactive AD, but its diversity and scale pose challenges for automated systems.
• Existing AD methods often require extensive engineering knowledge of ESD, limiting their applicability.

Purpose of the Study:

• To present a novel framework, Spectral and Time Autoencoder Learning for Anomaly Detection (STALAD), for automated in-line AD.
• To address the challenges of diverse and large-scale ESD in semiconductor manufacturing.
• To enable proactive yield and operations management through efficient AD.

Main Methods:

• Developed the STALAD framework incorporating four innovations: cycle series and spectral transformation (CSST) from ESD, unsupervised learning using Stacked AutoEncoders on CSST, hypothesis testing for AD, and dynamic procedure control for periodic and parallel learning/testing.
• Applied CSST to transform ESD for effective feature extraction.
• Utilized Stacked Autoencoders for unsupervised learning of normal operational patterns from transformed ESD.

Main Results:

• STALAD successfully learns normal operational patterns from ESD without requiring prior engineering knowledge.
• The framework demonstrates tolerance to some abnormal data during training, ensuring robustness.
• STALAD performs accurate anomaly detection and is efficient and adaptive for semiconductor fabrication applications.
• Applications on an HDP-CVD tool validated the framework's effectiveness.

Conclusions:

• STALAD offers a novel, automated approach to in-line anomaly detection in semiconductor manufacturing.
• The framework enhances proactive equipment maintenance and yield management by effectively utilizing ESD.
• STALAD complements existing methods like control wafer monitoring, potentially enabling earlier detection of equipment anomalies and their impact on process quality.