運用資料挖礦技術進行影響良率學習之因素分析 - 以某半導體廠製程為例

Using Data Mining Technology to Analyze the Effective Variables of Yield Learning - A Case Study of a Semiconductor Manufacturing Process

呂明澤(Ming - Chj Lu)

指導教授 ： 劉志明

國立清華大學/工學院/工業工程與工程管理學系/碩士(2007年)

https://doi.org/10.6843/NTHU.2007.00060

摘要

在半導體製造管理中，良率是反映製程是否有重大問題的一項績效指標，也會直接反映在晶圓廠的成本與獲利能力上。目標良率能否在一定時間內達到主要是取決於良率的學習率。影響良率學習的因素有很多，例如製程、設備與產品組合的複雜性、操作員或工程師的經驗、設備的新舊、製程的週期時間、製程導入的時間、在製品的數量等。然而，不同生產優序(production priority）的貨批（lot）其生產週期及產量與良率學習之間的關係卻很少被探討，因此如何釐清不同生產優序下的生產參數對於良率學習的影響，以及如何決定最適生產優序下的產品組合，以追求最低的成本支出和最大的良率學習，是本論文之研究重點。
首先本研究以生產週期時間等為主要影響變數，配合不同的生產優序，藉由類神經網路（artificial neural networks）結合田口方法（Taguchi Methods）的最適化模型為分析工具，搭配灰色系統理論（grey system theorem）的分析，來探討不同生產優序下的各種影響變數對良率學習的影響。並且在確認類神經網路之準確度後，將先前所得到的關係與影響度加以應用，以得到良率學習變數控管的定性建議與最適化生產規劃的定量分析。
透過本研究所發展出來的最適化類神經模型，可以協助釐清製程改善的因子並加以控管，也可以決定最適化生產優序批量組合與最適化產能擴充模式，以改善半導體廠整體的獲利。

關鍵字

良率學習 ； 生產優序 ； 生產週期時間 ； 生產量 ； 類神經網路 ； 灰色系統理論

並列摘要

In semiconductor manufacturing, yield learning is the most critical issue for process improvement. Speeding up the yield learning period can reduce the production cost and enhance the business profit. There are many papers which discussed the variables affecting yield（i.e. scheduling, dispatching, cycle time control, operator education…etc.）, but few studies have quantified the impact of the cycle time and the quantity on the yield learning rate based on the historical data of an existing wafer fabrication facility (fab).
In this study, we have constructed a hybrid model using Taguchi Methods and artificial neural networks to find the relationship between the yield learning rate and the related variables (cycle time, product quantity, and production priority) within different stages of a fab life cycle process. Furthermore, the grey system theorem is also applied in order to compare the accuracy and to avoid the drawback in the explanation of the hybrid model.
We can use the results from the analysis of the hybrid model to get insight into the impact of cycle time and product quantity of different production priority on the yield learning rate of a fab life cycle process.
Furthermore, by incorporating the hybrid model and cost function to evaluate the financial benefit of different combination of variable conditions, the optimal hot lot ratio and the optimal capacity ramp-up speed can be found.
Based on the result of the qualitative analysis and the quantitative application, it can reduce the time for process improvement and achieve an ideal yield learning rate for a fab.

並列關鍵字

Taguchi Methods ； Artificial Neural Networks ； Cycle Time ； Product Quantity ； Production Priority ； Grey System Theorem

參考文獻

1. Balasubramaniam, S., A. Sarwar, and D.M.H. Walker（1997）, “Yield Learning in Integrated Circuit Package Assembly” IEEE Transactions on Components, Packaging ,and Manufacturing Technology – Part C, Vol.20, No.2, Apr. 1997.

2. Bohn, R.E.（1995）, “Noise and Learning in Semiconductor Manufacturing” Management Science, Vol.41, No.1, pp.31-42, 1995.

3. Brian, C., R.C. Warner, B. Bidanda and L.N. Kim（2000）, “Predicting glass furnace output using statistical and neural computing methods”, International Journal Product & Research , Vol.38, No.6 , pp.1255-1269, 2000.

4. Cavalieri, S., P. Maccarroneb, and R. Pinto（2004）, Parametric vs. neural networkmodels for the estimation of production costs: A case study in the automotive industry” International Journal Product & Economic 91, pp.165–177, 2004.

7. Cunningham, S.P., and C.J. Spanos（1995）, “Semiconductor Yield Improvement: Results and Best Practices”, IEEE Transactions on Semiconductor Manufacturing Vol.8, No.2, May, 1995.