Abstract
In the age of digital information, large amounts of data are recorded in data warehouse or database. Such data may provide a rich resource for knowledge discovery from database and decision support. During the fabrication process, a large amount of process data will be automatically or semi-automatically recorded and accumulated in the engineering database for process monitoring, fault diagnosis and manufacturing management. However, in semiconductor industry, the yield of a silicon wafer is affected by many manufacturing factors that are often interrelated. It is difficult for domain engineers to find possible root causes rapidly and efficiently by own domain knowledge or rule of thumb. In this study, we construct a data mining conceptual framework for analyzing semiconductor manufacturing data, and propose a hybrid decision tree approach, including Kruskal-Wallis test, chi-square interaction detection, and variance reduction splitting criterion to explore the huge engineering data to analyze the semiconductor manufacturing data and infer the possible fault causes of manufacturing process variation. The information is helpful to engineers as the basis of the trouble shooting and defect diagnosis. In this study, we use real data to compare the performance of hybrid decision tree approach with current decision tree algorithms, and apply a real case from a semiconductor fabrication company as empirical study and the results show the practical viability of this approach.