隨著半導體生產技術的發展與進步，以及晶圓生產尺寸面積的增大，良率高低一直被半導體產業視為能否獲利的重要指標，因此良率預測模式已然成為半導體產業中相當重要的一個討論議題。早期的研究大多使用晶圓的缺陷數及晶圓面積等變數配合一固定的分配建構良率預測模式，但隨著製程的進步及晶圓面積的增加，晶圓缺陷的分佈有隨機及系統性因素所構成，因而加入了缺陷群聚指標來加以修正模型以提高預測準確度，不過此等固定分配之預測模式並不一定適合於每一種半導體製程，且有可能造成良率預測的不準確及不可靠。本研究利用晶圓允收測試(Wafer Acceptance Test, WAT)資料與晶圓針測良率(Circuit Probe Yield, CP)間之相關性，篩選WAT參數以減少輸入變數的數目，並藉由倒傳遞類神經網路(Backpropagation Neural Network, BPNN)、廣義迴歸類神經網路(General Regression Neural Network, GRNN)及集群資料處理技術(Group Method of Data Handling, GMDH)等分析方法，提出一簡便、有效率且準確度高的良率預測模式，以提供半導體業者於良率管理作業之參考。研究結果顯示使用倒傳遞類神經網路進行晶圓允收測試參數的選取，再配合集群資料處理技術建構良率模型，可達到有效縮減輸入變數、簡化預測模型的目標，且對良率的預測能夠有足夠的準確度及良好的預測效果。

Wafer die yield is a key index of business profit at the semiconductor industry as the development of the semiconductor production technology and wafer size increases. Thus, the wafer die yield prediction models become a very important issue to this industry. Early researches mostly use such parameters as defects on wafer and wafer size cooperate with a regular distribution, such as Poisson distribution or Negative Binomial distribution, to establish the wafer die yield prediction model. However, these prediction methods of yield models might not be suitable and accurate for all semiconductor manufacturers. In this study, we discover the connection between the wafer acceptance test (WAT) data and circuit probe yield. A simple, efficient and accurate wafer die yield prediction model with less WAT parameters is proposed. Various Neural Network analysis techniques such as Backpropagation Neural Network (BPNN), General Regression Neural Network (GRNN) and Group Method of Data Handling (GMDH) are analyzed and compared. Our Result shows Backpropagation Neural Network is valid and efficient in selecting WAT parameters. The GMDH is recommended for establishing the model because of its ability to catch data pattern with less variables and higher accuracy.