Improving Test and Diagnosis Efficiency through Ensemble Reduction and Learning

Machine learning is a powerful lever for developing, improving, and optimizing test methodologies to cope with the demand from the advanced nodes. Ensemble methods are a particular learning paradigm that uses multiple models to boost performance. In this work, ensemble reduction and learning is explored for integrated circuit test and diagnosis. For testing, the proposed method is able to reduce the number of system-level tests without incurring substantial increase in defect escapes or yield losses. Significant cost from test execution and set-up preparation can thereby be saved. Experiments are performed on two designs of commercially fabricated chips, for an overall population of >264,000 chips. The results demonstrate that our method is able to reduce 29.27% and 21.74% of the number of tests for the two chips, respectively, at the cost of very low defect escapes. For failure diagnosis, the framework is able to predict an adequate amount of test data necessary for accurate failure diagnosis. Experiments performed on five standard benchmarks demonstrate that our method outperforms a state-of-the-art work in terms of data-volume reduction. The proposed ensemble-based methodology creates opportunities for improving test and diagnosis efficiency.