半導體晶圓製造週期時間與其變異的縮短以及晶圓製造週期時間的差異化服務之研究

Developing Novel Approaches for Reduction of Mean and Variation of Manufacturing Cycle Time and Differentiation of Service for Manufacturing Cycle Time in Semiconductor Wafer Fabrication

郭仲仁

指導教授 ： 簡禎富

國立清華大學/工學院/工業工程與工程管理學系/博士(2010年)

摘要

Cycle time reduction is crucial for semiconductor wafer fabrication companies to maintain competitive advantages as the semiconductor industry is becoming more dynamic and changing faster. Nevertheless, while the number of mask layers has being increased in more advanced technology nodes, the theoretical cycle time is lengthened. Thus there is a need for novel approaches to continuously reduce cycle time for semiconductor wafer fabrication. Furthermore, as semiconductor devices are increasingly employed in consumer electronics, differentiated cycle time service through hot lots is crucial for semiconductor wafer foundries to enhance competitive advantages in the consumer era. However, more hot lots lead to longer mean and variance of cycle times. There is also a need for developing a model to simulate the impact of hot lots on the mean and variance of cycle times of both regular lots and hot lots so as to enable differentiated cycle time services.
On one hand, the existing queueing models for predicting the cycle times of tool sets in a semiconductor wafer fabrication facility (fab) have limitations in real settings, and simulation models are not feasible to deal with all the tool sets in a fab. On the other hand, little research has been done to predict the cycle times of tool sets with tool dedication and waiting time constraint. This study aims to fill the gap by developing data mining approaches to model the complex relationships between cycle times and related factors for individual tool sets taking into account tool dedication and waiting time constraint. The proposed data mining approach can help practitioners to explore manufacturing intelligence from the rich data in production database. In particular, several data mining techniques including neural networks and support vector regression are compared to investigate which one is most suitable for modeling the cycle times of tool sets. Moreover, this study concludes that classification and regression tree (CART) is more suitable than the competitive technique to rank the potential variables that might influence cycle times. Thus the relevant variables can be identified by combining CART and the selected technique for modeling the cycle times. By exploring the patterns from the built data mining models, the impact of hot lots on cycle times can be modeled so as to support the decision making for determining the quota of different classes of hot lots given a required lead time for regular lots. Furthermore, the data mining models can be utilized to prioritize related factors so as to effectively make use of the limited resources for reducing mean and standard deviation of cycle times. Moreover, this study proposed decision analysis models with regard to tool allocation among interchangeable tool sets and the determination of standard WIP for individual processing steps. The proposed decision analysis models can provide additional solutions for cycle time reduction. An empirical study was conducted in a wafer fab of a leading semiconductor company in Hsinchu, Taiwan to validate the proposed approach. The results showed practical viability and significant effectiveness of the proposed approaches.

關鍵字

縮短製造週期時間 ； 製造週期時間差異化服務 ； 資料挖礦 ； 半導體晶圓製造 ； 限定機台 ； 等候時間限制 ； 決策分析 ； 製造智慧

並列摘要

半導體產業變化愈來愈快，其產業脈動也益發難以掌握，半導體晶圓製造業者必須不斷縮短製造週期時間以維持其競爭力。然而，光罩層數隨著先進半導體製程的演進而持續增加，使得製造週期時間愈來愈長，因此，半導體晶圓製造業需要更多新的方法來維持或縮短製造週期時間。此外，電子業已逐漸邁入消費性產品的時代，客戶對製造週期時間較短的急單需求也愈來愈強烈，晶圓代工業者必須有能力提供製造週期時間的差異化服務，並且提供更多的急單服務，方能在消費性電子時代中取得競爭優勢。但較多的急單會拉長製造週期時間的平均值與變異，要提供製造週期時間的差異化服務，必須建立一個模式來衡量急單對製造週期時間平均值與變異的影響。
本研究目標係發展資料挖礦的相關技術來建立製造週期時間與相關因素之間的關聯模式，並考量限定機台與等候時間限制這兩個晶圓廠中普遍存在的限制因素。在真實的半導體晶圓廠環境中，運用等候模式來預測製造週期時間有其限制，而運用電腦模擬來建立一個晶圓廠中所有機台的製造週期時間模式，在實務上的可行性也很低。晶圓製造流程中累積了豐富的製造資料，本研究運用資料挖礦技術來挖掘存在於製造資料中的製造智慧，以協助晶圓廠找到改善製造週期時間的機會。本研究比較許多類神經網路以及支援向量迴歸技術，確認當中最適合用來預測機台製造週期時間的方法。同時，本研究也評估了兩種對可能影響機台製造週期時間的相關因素重要度加以排序的方法，以確認各機台相關的影響因素。運用本研究所建立的資料挖礦模式，急單對於製造週期時間的影響可以被量化，進而提升製造週期時間差異化服務的可行性。運用相同的資料挖礦模式，可以針對相關影響因素做重要性排序，協助晶圓廠運用有限資源來改善重要的影響因素，以發揮縮短交期的最大效果。此外，本研究進一步建立相關決策分析模式來協助晶圓廠針對可相互支援的機台群組做最佳化的機台數配置，以及決定各加工站的最適在製品存貨水準，這些決策分析模式可進一步提供縮短製造週期時間的機會。本研究所提出的方法已被實際應用於新竹科學園區某晶圓製造廠，應用的結果驗證了本研究所提出的方法在實務上的可行性，並且顯示出可觀的效果。

並列關鍵字

Cycle time reduction ； differentiated cycle time service ； data mining ； semiconductor wafer fabrication ； tool dedication ； waiting time constraint ； decision analysis ； manufacturing intelligence

參考文獻

Akçali, E., Nemoto, K., and Uzsoy, R. M., 2001, “Cycle-time improvements for photolithography process in semiconductor manufacturing,” IEEE Transactions on Semiconductor Manufacturing, vol. 14, no. 1, pp. 48–56.

Atherton, R.W. and Dayhoff, J. E., 1986, “Signature analysis: simulation of inventory, cycle time, and throughput trade-offs in wafer fabrication,” IEEE Transactions on Components, Hybrids, and Manufacturing Technology, vol. 9, no. 4, pp. 498–907.

Backus, P., Janakiram, M., Mowzoon, S., Runger, G. C., and Bhargava, A., 2006, “Factory cycle-time prediction with a data-mining approach,” IEEE Transactions on Semiconductor Manufacturing, vol. 19, no. 2, pp. 252–258.

Baek, D.H., Yoon, W. C., and Park, S. C., 1998, “A spatial rule adaptation procedure for reliable production control in a wafer fabrication system,” International Journal of Production Research, vol. 36, no. 6, pp. 1475–1491.

Blum, A. and Langley, P., 1997, “Selection of relevant features and examples in machine learning,” Artificial Intelligence, vol. 97, no. 1–2, pp. 245–271.