Application of 1D ResNet for Multivariate Fault Detection on Semiconductor Manufacturing Equipment

Area of Science:

• Manufacturing Process Optimization
• Artificial Intelligence in Industry
• Semiconductor Manufacturing

Background:

• Reducing manufacturing costs and enhancing productivity are key objectives.
• Massive amounts of sensor data are generated during manufacturing, posing challenges for efficient utilization.
• Fault Detection and Classification (FDC) is crucial for evaluating production machinery conditions.

Purpose of the Study:

• To address the challenge of identifying defective wafers in semiconductor manufacturing.
• To propose a multivariate fault detection method using raw time-series data from multiple sensors.
• To analyze and characterize wafer status using features from specified process tools.

Main Methods:

• Implementation of a 1D Residual Network (ResNet) algorithm for multivariate fault detection.
• Analysis of raw time-series data collected from sensors throughout the semiconductor manufacturing process.
• Addressing the gradient vanishing problem in deep neural networks, which affects plain 1D Convolutional Neural Networks (CNNs).

Main Results:

• The 1D ResNet method effectively identifies anomalous wafers by analyzing multivariate time-series data.
• Experimental results demonstrate superior performance and accuracy compared to plain 1D CNN-based methods.
• The proposed method overcomes limitations of deeper plain 1D CNNs, such as the gradient vanishing problem.

Conclusions:

• The 1D ResNet-based fault detection method is effective and accurate for identifying abnormal wafers in semiconductor manufacturing.
• This approach offers a robust solution for leveraging large-scale sensor data in complex manufacturing environments.
• The study highlights the potential of deep learning, specifically ResNet, for improving quality control in semiconductor production.