Statistical Analysis and Design for Semiconductor Production System

• Main Authors: Puffy Lin, 林悅慈
• Other Authors: Argon C. K. Chen
• Format: Others
• Language: zh-TW
• Published: 2000
• Online Access: http://ndltd.ncl.edu.tw/handle/32954053864977012200

碩士 === 國立臺灣大學 === 工業工程學研究所 === 88 === Wafer fabrication, one of the key processes in semiconductor manufacturing, is perhaps the most complex manufacturing process. A major process flow in a wafer fabrication factory may contain over 200~300 separate steps or operations. It’s complicated processes (reentrant characteristics, queue time limit, batching requirement, etc.) and uncertainties (machine down, rework, yield, etc.) make its production planning and control even harder. In this research, we demonstrate how to make quality decisions for both long-term manufacturing strategy and short-term manufacturing control using simple, but effective, statistical modeling and analysis methods. The first objective of this research is to develop a statistical procedure to select manufacturing strategies that optimize production performance. By conducting full factorial experiments with a validated simulation model, the simulation results are analyzed using off-line quality engineering techniques including Taguchi method and response surface method (RSM). The affecting factors selected in the experiments include product mix, hot lot ratio and batching rule. And production performance measures include throughput, WIP, and cycle time. Experimental data are collected and analyzed to determine factor levels that optimize the production performance through S/N ratios and/or steepest ascent methodologies. Based on the experimental results, we’ll be able to propose the best manufacturing strategies to manage the trade-off among different or conflicting production objectives and to reduce the production variability. The second objective is to build statistical time series models for real-time control and monitoring. We first propose a functional autoregressive moving average (F-ARMA) time series model to characterize the experimental production data. The F-ARMA model validates the feasibility of time series applications to semiconductor manufacturing systems. To illustrate how to apply on-line quality monitoring and control techniques to production systems, we collect actual production data from a local semiconductor fabrication factory and use time series models, including ARIMA, transfer function, and VARIMA models, to characterize the factory operations. Two control charts are then constructed: common-cause and special-cause control charts. The common-cause chart shows the trends of production conditions, which are accurately predicted by the time series models, and is used for real-time manufacturing control. The prediction residuals are used to construct the special-cause control chart for monitoring and detecting any production deficiencies caused by unusual operational problems.