Graded structure of defect categories in automated defect classification

Date

1996-05

Journal Title

Journal ISSN

Volume Title

Publisher

Texas Tech University

Abstract

Defect review and classification are time consuming, monotonous, fatiguing, and very subjective tasks. Large amount of variability has been observed from different operators, same operator for the same wafer, from wafer to wafer, or from day to day depending on the type of defects observed. This research hypothesizes the problem as due mainly to the graded structure of semiconductor defect categories. Every category exhibits a graded structure. Graded structure refers to degree of membership representation from the most typical to atypical members of a category to those nonmembers that are least similar to the category members.

Two levels of categorization and therefore two levels of graded structures occur in knowledge-based defect classification which can adversely affect the performance of the system. The first level of graded structure occurs in the design of defect knowledge base when the defect expert describes defect categories in terms of defects' visual features using linguistic variables. Feature values which are acquired subjectively and represented in natural language are most subjected to the limitation associated with graded structure. Graded structure level two occurs in the actual defect classification where feature similarities between the actual defect and the categorical defects are compared.

The focus of this research is minimization of the effect of graded structure in automated defect classification for patterned semiconductor wafers. This research provides insights into the use of standardization and feature combination as a way to minimize the effect of graded structure on categorization. Level one graded structure can be minimized by standardizing defect features, e.g. by representing them as fuzzy sets. The level two graded structure can be minimized by combining human-based and computationally-derived defect features in categorization. Human generated object features are directly visible and recognizable to a human observer while computationally-derived features are not directly perceptible by humans.

Description

Citation