Improving Fault Detection Accuracy in Semiconductor Manufacturing with Machine Learning Approaches

Abstract

The production of semiconductors is among the most technologically complex and intricate industrial processes. Given the hundreds of steps involved in semiconductor production, a defective wafer detection system that enables early wafer identification is necessary. Virtual metrology (VM) and statistical process control (SPC) have been used to identify defective wafers. The purpose of this research is to examine how machine learning (ML) methods may be used to enhance the accuracy of semiconductor manufacturing defect detection. By leveraging the WM811K dataset, which includes over 800,000 wafer images with multiple defect categories, the research applies a Convolutional Neural Network (CNN) integrated with data augmentation to enhance model performance. The proposed CNN-AUG model effectively addresses challenges such as data imbalance and overfitting, yielding an accuracy 98.56%, precision 98.77%, recall 98.78%, and an F1-score 98.77%. Comparative analysis with VGG19 and XGBoost demonstrates the superior performance of CNN-AUG in capturing intricate spatial features and improving fault detection efficiency. The results highlight the potential of ML-based approaches for optimizing semiconductor manufacturing processes, reducing defects, and enhancing yield.