積體電路良率模式之修正與研究

Abstract

隨著積體電路複雜度的提高、晶片面積的增大以及缺陷群聚的出現，原有的良率模式如Poisson model等，逐漸發生嚴重的偏差，使得其在良率分析上的效果大打折扣。即使是現今廣為積體電路製造業使用的的Negative binominal model，亦在實証研究中，遭到多位學者的質疑，例如其中的缺陷群聚參數與晶片面積相關，且可能逐批不同等等。 本論文的目的，即是針對現行良率模式在應用上的缺失，且考慮缺陷群聚的因素，透過統計分析的技巧，找出一個預估能力較佳的良率模式。本論文首先將quadrat analysis、hierarchicalclustering以及correlation coefficient method三者組合成一個群聚分析程序，而能有效地從wafer map上尋找出缺陷群聚的位置、群聚所含成員和群聚間的相對強度；然後再結合晶圓分區的觀念而發展出一套關於Poisson model的修正流程，以消除缺陷群聚現象對Poisson model的影響。 而透過實例的分析發現，本論文所提之群聚分析程序確能有效的界定缺陷群聚現象，且修正後之Poisson model其良率預測的準確性優於目前較常使用的良率模式。此外，由分析的過程中也可知，良率分析實可應用於生產現場以及高強度的群聚現象可能源於機器異常或人為疏失之故。

As the complexity of integrated circuits rise, the size of chip increases and the clustering of defects emerges, existing yield models, such as the Poisson model, gradually result in deviation and thus become impractical. Even the negative binominal model which are most popular in the IC industry nowadays since its defect-clustering parameter is related to the chip size and probably may vary with the lot are questioned by many scholars. The objective of this research is to aim at the drawbacks of present yield models, take defect clustering in account, and then, through statistical analysis, develop a yield model with better yield prediction capability. In this resrearch, we first combine quadrat analysis, hierarchical clustering method, and correlation coefficient method as a defect-clustering analysis procedure to locate defect clusters on wafer maps, identify members of clusters and the relative strength of clustering. Then adopting the concept of wafer partition, we develop a procedure to modify Poisson model with a view to effacing the impact of defect clustering on yield prediction of the poisson yield model. Through experimental analysis, the clustering analysis procedure proposed by this research are verified to be good at detecting defect clusters, moreover the yield predicton ofmodified Poisson model is more accurate than existing, yield models. Besides, we conclude that yield analysis is applicable in production line, and the strong defect-clustering is likely resulted from mechanical malfunction or human errors.