Automatic Timeiseries Model Generation for Real - Time Statistical Process

To my father, my mother and Stephanie 1 for their love and support. Acknowledgment I I would like to express my deep appreciation and gratitude to my research advisor, Dr. Costas J. Spanos, for his support and guidance through my graduate studies an the latter 9 portion of my undergraduate studies. I also thank Dr. John A. Rice for being project report committee and for his valuable insights in time-series modeling. Special thanks are due to two colleagues in the BCAM group: Mr. Eddie work in developing the BCAM Real-Time SPC interface, and Ms. Sherry her knowledge and expertise and providing insightful feedbacks and help and friendship are greatly appreciated. I would also like to extend my gratitude to the rest of the BCAM group for making my graduate experience an enjoyable one. They are Mr. ABSTRACT As integrated circui; designs become more complex, in compliance with Mdore's h w , assuring the production quality of these complex integrated circuits becomes increasingly difficult. Consequently, semiconductor manufacturers must focus on achieving tighter real-time process control in order to obtain justifiable production yields as well as sustain profitability in an increasingly competitive marketplace. Traditionally, equipment and process faults are being discovered by " in-line " measurements done between process steps. However, due to an increased pressure to produce of a highly diverse product mixture in shorter cycle times, equipment and process faults must be detected in real-time. However, because real-time process contrbl requires the analysis of real-time equipment sensor data, traditional statistical process control (SPC) techniques [ 13 cannot be readily applied to the sensor data due to their non-stationary, auto-correlated and cross-correlated characteristics. The Berkeley Computer-Aided Manufacturing (BCAM) Real-Time SplC system utilizes econometric time-series models [2] in order to filter real-time readiqgs of any existing autocorrelations. In addition, multivariate statistics, in particular, the Hotelling's TL statistic [3], are then used in order to combine the various cross-correlated signals into a single statistical score. This T2 statistic is monitored with a single-sided contr real-time SPC [4]. The objective of this project was to develop and implement an algorithm automating the time-series model generation process for real-time SPC. Furthermore, modifications must be made to the real-time SPC scheme in order to accomodate the batch nature of single-wafer processing operations. As a result, the BCAM Real-Time SPC scheme has been modified, and an automatic time-series model generator has beem developed. The model generator has …