應用於半導體製程資料分析之正規化機器學習方法

Abstract

在晶片製造的過程中，各個製作階段或步驟皆會記錄下大量的資訊。而良率則透過這些資料得以提升，例如：某參數數值的漂移可能即是造成錯誤的原因，或多個參數的交互作用也可能影響良率。但是，大量的資訊會造成分析的困難。 這份研究致力於利用正規化機器學習方法來降低資料維度，並辨識出影響良率的關鍵參數。這份報告的資料分別由兩家在台灣的半導體公司提供，使用套索回歸和懲罰式支持向量機來建立模型。 實驗結果顯示，參數的確會影響良率表現，而提出的方法可以有效降低資料維度並找出關鍵因子。此外，懲罰式支持向量機在高維度的實作方法也會在此論文中介紹。

During wafer fabrication, volumes of data were recorded from monitoring through multi-stage and multi-step of manufacturing process. Yield quality is expected to be increased by discovering the information from the recorded data. For example, a drift in the value may cause a failure, or a certain combination of parameters can lead to better wafer performance. However, the great volume of recorded data may lead to difficulties in identifying the data. To identify the root causes of defects and process parameters that are linked to yield performance, this study aims to develop a regularized machine learning method to reduce the dimension of parameters and indicate the key features which lead to the lower quality in yield. Data for this research were obtained from two semiconductor foundry companies in Taiwan. LASSO regression and penalty SVM are introduced to model the data in this thesis. The results show that the drop of yield quality is truly affected by some parameters and the regularized method is effective in shrinking the model to the root causes. Besides, a penalty SVM in high dimension implementation is introduced and realized in this work.