文摘
As integrated circuit technology continues to scale, new and complex failure mechanisms occur forcing test to evolve from its traditional role of pass/fail screening to adaptive data-driven methods capable of diagnosing and predicting failures, modeling highly nonlinear processes, and coping with variability. Detecting and understanding failures associated with problems such as atypical test results, systematic shifts, lithographic variability, test escapes, and high volumes of noisy parametric measurements requires a holistic perspective for informed decision making. A new test paradigm must leverage both design and test data to uncover valuable knowledge that can explain and correct complex failure mechanisms while offering better tradeoffs between test cost and quality. Such an approach represents a shift from simply uncovering statistical significance in test data to a method that explains statistical findings through interpretable and actionable knowledge. This dissertation introduces a knowledge discovery process for test that eases the transition from well-established pass/fail methods to newly emerging statistical testing and modeling. A knowledge discovery approach is used to develop several automatic process flows for analyzing design and test data, enabling the extraction of both descriptive and predictive knowledge, beginning with raw data and ending with valuable insight. A variety of preprocessing, data mining, and visualization methods are used to build a unified knowledge discovery framework with several applications in test. Silicon measurements and design data from four industry products currently in production are analyzed to show how a knowledge discovery process can perform data driven failure diagnosis, outlier based failure screening, predict variability in nanoscale lithography, automate non-parametric behavioral modeling, and reduce test cost by parametric test set optimization.