應用線性迴歸模型於射頻IC特性預測

Abstract

製程控制是半導體業很重要的一個環節。為了預測射頻IC特性，目前多半以人工方式挑選變數，但此法不能同時監控多個參數以利晶圓挑選，故易導致良率偏低，造成損失。本論文以統計學之線性迴歸分析作為預測射頻IC特性模型。本研究以現有晶圓製程資料作為輸入變數，並以最終產品的測試結果作為輸出變數，透過有效的變數選擇程序，即向前選擇法，並利用F檢定準則篩選出重要的輸入變數，以建立準確且強健的預測模型。最後，我們利用測試資料來驗證預測模型的實際效用，結果顯示準確率可達92.5%以上，相較於原本的人工方式，良率因此獲得大幅改善。

Process control is a critical factor in semiconductor manufacturing. To predict IC characteristics, the present industry still relies heavily on human expertise. Manual selection for explanatory variables cannot deal with multi-variables effectively and thus often results in poor prediction accuracy. In this thesis, we adopt multivariate linear regression to predict RF IC characteristics. The wafer fabrication data provided by IC manufacturing company are taken as input variables, and the final function test data are target outputs. By adopting variable selection scheme, we are able to identify key input variables so as to enhance the model accuracy without losing robustness. Our study showed that the proposed approach can achieve 92.5% accuracy for the validation data. The yield rate was improved significantly.