Deep Learning Based Recognition of Mixed Defect Patterns Generated in Semiconductor Fabrication Process

  • Ghalia Tello

Student thesis: Master's Thesis


Ghalia Tello, 'Deep Learning Based Recognition of Mixed Defect Patterns Generated in Semiconductor Fabrication Process', M.Sc. Thesis, M.Sc. in Electrical and Computer Engineering, Department of Electrical and Computer Engineering, Khalifa University of Science, Technology and Research, United Arab Emirates, June 2016. Semiconductor manufacturers are concerned with producing defect-free wafers to maximize yield and minimize cost. Tracing back and fixing the source of the defects can save the companies from enormous losses. Fortunately, these defects tend to cluster and form spatial patterns that provide useful information to identify potential problems during the fabrication process. Machine learning approaches have been utilized to classify these defects in order to help locate and fix their root cause. This thesis extends existing literature by proposing a Deep Structured Learning (DSL) system to identify and classify 'single' and 'mixed' defect patterns. Single patterns appear because of faults that occur from a single problem in manufacturing whereas mixed defects contain multiple patterns. The proposed DSL system first applies a spatial filter to remove random noise and is followed by three classification components. The first classification stage splits the dataset into single and mixed defect patterns by using a Decision Tree (DT) based rules that was trained using a feature selection technique called Backward Elimination Ranking (BER). Defects with single patterns are then classified using a shallow Randomized General Regression Network (RGRN), while the mixed patterns are classified using the deep learning approach of Deep Structured Convolutional Network (DSCN). This combination aims to give a complete system that can classify a wider range of defect patterns and provide optimal performance. We have shown that our proposed model achieved a total accuracy of 86.17% in classifying different types of single and mixed defect patterns that are commonly found in real scenarios, outperforming other stateof- the-art machine learning models reported in the literature. Index Terms—deep learning, convolutional neural networks, feature selection, semiconductor wafers defects, mixed patterns
Date of AwardJun 2016
Original languageAmerican English
SupervisorYousof Al Hammadi (Supervisor)


  • Deep Learning
  • Convolutional Neural Networks
  • Feature Selection
  • Semiconductor Wafers Defects
  • Mixed Patterns.

Cite this