晶圓研磨之平坦度預測模型研究

Abstract

化學機械平坦化(Chemical Mechanical Planarization, CMP)是使晶圓表面平坦化的方法之一，是IBM公司於1985年發展CMOS產品時研發成功的一項技術，其目的是應用研磨技術將晶圓上凹凸起伏的介電層或金屬層平坦化；平坦化的過程除了利用機械應力作用外，並結合研磨液在研磨介面間產生的化學作用，經由研磨運動而達到平坦化的目的。此技術為目前應用於半導體晶圓片表面平整化的尖端技術。 雖然CMP製程在半導體廠的地位日漸重要，因為其所涉及的物理、化學、機械以及研磨機台、研磨液、研磨墊等設備與耗材之間相互作用的影響十分複雜，故各大晶圓廠均致力於提升對CMP製程的研發，以期提升製程的控制能力。本研究利用生產線上的實驗數據，以研磨墊物理性質為實驗變數，找出研磨墊物理特性對研磨後晶圓表面平坦度的影響，並依此建立迴歸預測模型，以協助工程師選擇合適的研磨墊達成平坦度的規格要求。 研究結果顯示研磨墊主要物理特性中，厚度、硬度與撓曲度對研磨後晶圓表面平坦度的影響較大，且分別呈負相關、正相關與負相關，可供工程師日後選擇研磨墊時參考；而依研磨墊物理性質所建立的多元迴歸模型與灰關聯迴歸模型在預測的效果上均達到相當高的準確度；然而因實驗數據較少，以及預測變數間存在相關性的原因，因此研究結果顯示使用灰關聯的預測模型較能解釋及預測研磨後晶圓表面平坦度的關係，主要原因為在資訊不充足的系統環境中，灰關聯分析較能分析出系統中各因素間主要的關係，並找出影響目標值的重要因素，從而獲取較多的資訊並作出較準確的預測。

Selective oxide CMP methods like ILD and PMD CMP are commonly used in achieving global planarization for schemes of quarter micron processes and beyond. PMD & ILD CMP process offers great advantages in processing devices with better global planarity, enabling the use of additional metal layers on a chip manufacturing, tighter dimensional control in lithography, and enhancing yield by improving planarity to reduce the probability of metal stringers after metal damascene polishing. In PMD planarization using CMP method, the post CMP planarity is extremely sensitive to the pad properties and polishing tools employed. Polyurethane pads are commonly used for dielectric CMP application. Almost all the polishing performances like removal rate, WIWNU, WIDNU, defectivity, and planarity are strongly correlated with polishing pad properties. In this paper, we performed the statistic analytical method on the CMP behavior and try to set up a model to predict the post CMP planarity. The regression models which correlate polishing pad physical properties with post CMP wafer planarity have been explored for PMD CMP by using traditional Statistic Correlation and Grey Correlation analysis in this paper. As the polishing data quantity is limited, we find that Grey theory performs better in linear regression variable decision. By the step height regression model, we can get a clear picture about the influence of pad properties on post CMP step height. We find that stacked pad deflection and hardness are two major pad properties controlling the post CMP planarity. By examining the correlation of stacked hardness and deflection, these two properties are not totally independent to each other and the correlation is not low. This explains the model built with deflection and stacked thickness yield better fitness than the model built with stacked hardness and deflection both on Mirra and 6DS polishers. In conclusion, regression models derived from Grey Correlation analysis offer excellent fitting on post CMP planarity in this study.