運用資料探勘技術於半導體製程異常機台分析

Abstract

半導體製造程序包含數百個執行步驟，如何有效地判定異常機台其實是一個複雜的問題。本研究於每晶圓批完成探針測試之後即對其產生的晶圓圖進行分析，主要採用凝聚式階層分群演算法，並整合失效晶粒之空間分布特性、良率以及時間維度參數，藉此判斷發生異常的機台與反應室。在階層式分群演算法中，本研究評估以最小距離、最大距離、均值距離與平均距離之分群結果，歸納平均距離的分群表現最佳；另實驗結果同時證明本研究提出之系統化方法，不僅能夠有效率地分析晶圓與反應室層級的資料，同時也兼具客觀性和準確性。最後經由分群方法之比較，也證實階層式分群方法的效果優於K-means與EM。

There are hundreds of processes in semiconductor manufacturing. It is a difficult task to find abnormal equipments. This research constructs a model to analyze wafer bin maps after circuit probe testing. We have developed a hybrid approach that integrates hierarchical clustering, spatial characteristic, yield, and time dimension for abnormal equipments and chambers detection. We apply four hierarchical clustering algorithms, include single-linkage, complete-linkage, mean and average method, and observe that average clustering algorithm works the best with wafers from a semiconductor manufacturing company. The experimental results show that the proposed model can identify the faulty wafers and at chamber level efficiently and effectively. The results also indicated hierarchical clustering algorithm would perform better than K-means and EM algorithm.