應用改良型巨集式啟發式方法於TFT-LCD良率控制最佳化之研究

Abstract

TFT-LCD的市場在最近幾年成長得相當快速。每一個LCD廠商都想要增加自己的市場佔有率，所以良率控制對廠商而言是一個提升競爭力與市場佔有率的重要課題。在TFT-LCD的液晶面板組裝（LC Cell Assembly）製程中，主要的步驟是將薄膜電晶體面板（TFT）與彩色濾光片（CF）透過機械對位壓合，再注入液晶，組合成LCD面板。此製程難度甚高，是LCD面板製作程序中良率最低的一環。在壓合前TFT與CF面板上都會被切割成多個panel1。然而在TFT與彩色濾光片壓合的時候，兩者中只要有一者的panel是不良品，則壓出來的LCD panel也是不良品。如果使用隨機的方式來組合TFT與CF兩種面板，可能會造成很多不良品，使良率降低。另外，在製作LCD面板時，會使用分類器（sorter）來提升良率。在本研究中，我們首先使用兩種常用的巨集式啟發式方法（meta-heuristic method）：基因演算法（GA）以及模擬退火法（SA）來解決TFT與CF的組合問題。但是如果需要配對的cassette2數量太多，會造成可行解的數量太過龐大。本研究也提出兩階段基因演算法以及模擬退火法來有效率地產生品質較好的初始族群。在階段一中，我們利用TFT cassette與CF cassette一對一的情況下之最佳解來產生品質較好的初始解。在階段二中，我們就使用階段一的初始解來執行基因演算法和模擬退火法。研究結果顯示，兩階段的GA和SA其表現都優於一般的基因演算法和模擬退火法。

1 panel為液晶面板的最小單位 2 cassette為裝置TFT和CF面板的容器

Yield control plays an important role in the TFT-LCD manufacturing firms, and the post-mapping operation is a crucial step. The post-mapping operation combines one TFT plate and one CF plate to form a LCD. Each TFT and CF plate is divided into a number of panels. The LCD panel is acceptable only when both TFT and CF panels are good. The TFT-LCD manufacturing firms use the sorter, a kind of robot, to increase the yield for matching TFT and CF plates. Evidently, there will be a great loss if a random mapping policy is executed. In this study, we first apply two of the most popular meta-heuristic methods to solve the post-mapping problem: GA and SA. However, when the number of matched cassettes is large, the number of ways for choosing different matched objects will become so enormous that the initial population in GA (or initial solution in SA) should be selected with a proper procedure. That is, we propose a two-phased GA and SA to improve the performance of the initial population. The basic concept of phase one is to generate an efficient initial population (or initial solution). In phase one, the initial population is created based on the optimal solution to the cassette-matching problem. In phase two, we perform GA (or SA) with the initial population created in phase one. The four different heuristic algorithms are tested for the same data to compare the various ports in the post-mapping yield control problem. The result shows that proposed two-phased algorithms provide a more excellent solution than GA and SA.