Early reliability assessment by using deep censoring

Harry A. Schafft, Linda M. Head, Jason Gill, Timothy D. Sullivan

Research output: Contribution to journalArticlepeer-review

1 Scopus citations


A method is described for making direct characterizations of the early part of the intrinsic electromigration fail-time distribution of interconnects. The method involves stressing a large number of test lines only long enough for a relatively few lines to fail, enough to characterize the percentile of interest. Groups of test lines are electrically monitored to detect failures without having to measure individually the many lines on test. Two types of deep censoring (DC) are described: DC without removals (where more than one line failure in a group can be detected with confidence) and DC with removals (where, when one failure occurs, the other lines in the group are removed from the test). Sample estimates of sigma and of one or more early percentiles of the distribution are corrected for bias and their confidence limits calculated. DC offers important benefits over the present practice of placing tens of test lines on test to obtain sample estimates of t50 and σ that are used to extrapolate to the early part of the loge(fail-time) distribution. The benefits are reduced testing times, better confidence of the sample estimates of early percentages of the distribution, and the ability to detect extrinsic fail-time populations. A detailed procedure to implement the method is provided in the appendix. Published by Elsevier Science Ltd.

Original languageEnglish (US)
Pages (from-to)1-16
Number of pages16
JournalMicroelectronics Reliability
Issue number1
StatePublished - Jan 2003

All Science Journal Classification (ASJC) codes

  • Electronic, Optical and Magnetic Materials
  • Condensed Matter Physics
  • Safety, Risk, Reliability and Quality
  • Surfaces, Coatings and Films
  • Atomic and Molecular Physics, and Optics
  • Electrical and Electronic Engineering


Dive into the research topics of 'Early reliability assessment by using deep censoring'. Together they form a unique fingerprint.

Cite this