利用統計方法自動依工程忍受度判斷機台差異及其在半導體製程改善之應用

Abstract

在半導體產業, 機台表現之比較是提升量率的關鍵之一, 我們發展一套新的統計方法, TCP.不但可以適用在機台使用率不相同造成之樣本數不同之現象亦可主動依工程師設定之工程忍受度將機台分群,可大幅縮短工程師分析機台表現之時間,我們將提供統計模擬來說明我們方法之優點也提供兩個半導體業界之實例說明我們的方法可適用在量率提升與製程能力提升

In the semiconductor industry, tool comparison is a key task in the yield and the product quality enhancements. We developed a new method, called tolerance control partitioning (TCP), to automatically partition tools into several homogenous groups based on the related metrology results. This methodology is based on a hierarchical normal model and the implementation is carried out using a Bayesian approach. There are several advantages of using the TCP method. First, it takes into account the unbalanced usage of the tools in the manufacturing processes. Moreover, the “engineer’s tolerance control” can be incorporated into the TCP method via the specification of the priors in the Bayesian analysis, which justifies the significant difference between groups according to the experts’ knowledge. This specification not only has the advantage of adjusting the number of partition groups but also avoids the problem of having too many partition groups with small differences which is often encountered in the conventional approaches. Some simulation results illustrate the advantages of the TCP method compared to the method of classification and regression trees (CART). Moreover, the TCP method is applied to two real examples for the yield and Cp/Cpk enhancement in the semiconductor industry. Both results confirm the practical usefulness of the proposed method. For general applications, the TCP method is also useful for other similar problems such as the comparisons between several experimental recipes or the comparisons between different materials.