Statistical and Data Mining Methods for Test-Based Yield Learning

Yield losses causes are usually represented as a mix of factors including process systematic problems, random defectivity and systematic marginalities. The first two components were known as dominant effects in earlier technology nodes, and have been faced and discussed widely in the past. Systematic marginalities instead are often represented as the overlap of two sliding windows. The first one is known as the "process window" which is subject to "natural" process variations or spreads. The second is instead ruled by the product margins and performances. In this context, finding significant correlations is the key problem toward yield learning. The yield learning process works on multiple data sets with potentially huge data volumes, sometimes in presence of incomplete information. Statistical and Data Mining methods are known to be able to effectively deal with this kind of problems. The proposed embedded tutorial aims at providing an overview of some statistical and data-mining techniques applied to yield learning. Among the problems that can be solved we will discuss feature extraction (via Projection Pursuit, Principal Component Analysis, and Independent Component Analysis), classification (via clustering and unsupervised Neural Networks), handling incomplete data (using Maximum Likelihood and Expectation Maximization methods), and fault diagnosis (via categorical data analysis and Bayesian Networks). Several case studies will be used to introduce problems and methods. The analysis of Iddq signatures illustrates the problem of incomplete data and its solution via the Expectation Maximization method. A further example of incomplete data arises in the post-processing of ATPG diagnoses aimed at reducing the number of fault candidates associated with each observed defect. The automatic classification of Sort Maps can be addressed by means of unsupervised clustering techniques either statistical or neural. Finally, according to the ACID (AutomaticClassification/Interactive-Diagnosis) paradigm, process diagnosis builds on Sort Map classification by correlating its results with lot history information. Within this paradigm, an effective approach to the isolation of faulty equipments relies on categorical data analysis techniques. Presentations