Generative Adversarial Network-Based Fault Detection in Semiconductor Equipment with Class-Imbalanced Data

Area of Science:

• Semiconductor Manufacturing
• Process Control
• Data Science

Background:

• Fault detection in plasma etching is crucial for semiconductor manufacturing.
• Class imbalance in fault detection datasets hinders model training and performance.
• Optical emission spectroscopy (OES) is used for monitoring plasma etching processes.

Purpose of the Study:

• To apply Generative Adversarial Networks (GANs) to mitigate class imbalance in plasma etching fault detection.
• To improve the accuracy of fault classification in semiconductor processes.
• To enhance the effectiveness of fault detection algorithms using augmented datasets.

Main Methods:

• Acquired two datasets from silicon trench plasma etching using OES, simulating Mass Flow Controller (MFC) faults.
• Utilized GANs to generate synthetic data for minority classes in imbalanced datasets.
• Evaluated five existing fault detection algorithms on augmented datasets.

Main Results:

• GAN-generated data successfully compensated for class imbalance in OES datasets.
• Fault detection algorithms showed improved modeling performance with augmented data.
• The proposed GAN-based data augmentation enhanced fault classification accuracy.

Conclusions:

• GANs are a beneficial tool for addressing class imbalance in OES datasets for semiconductor fault detection.
• Augmenting minority class data with GANs improves the reliability of fault detection systems.
• This research contributes to more robust advanced process control in semiconductor manufacturing.